Received: 9 February 2022 Revised: 17 August 2022 Accepted: 5 October 2023

DOI: 10.1002/widm.1520

ADVANCED REVIEW

Deepfake detection using deep learning methods:

A systematic and comprehensive review

Arash Heidari 1 | Nima Jafari Navimipour 2,3 | Hasan Dag 4 | Mehmet Unal 5

1
Department of Software Engineering,
Haliç University, Istanbul, Turkey Abstract
2
Department of Computer Engineering, Deep Learning (DL) has been effectively utilized in various complicated chal-
Kadir Has Universitesi, Istanbul, Turkey lenges in healthcare, industry, and academia for various purposes, including thy-
3
Future Technology Research Center, roid diagnosis, lung nodule recognition, computer vision, large data analytics,
National Yunlin University of Science and
Technology, Douliou, Taiwan
and human-level control. Nevertheless, developments in digital technology have
4
Management Information Systems, Kadir been used to produce software that poses a threat to democracy, national security,
Has Universitesi, Istanbul, Turkey and confidentiality. Deepfake is one of those DL-powered apps that has lately sur-
5
Department of Computer Engineering, faced. So, deepfake systems can create fake images primarily by replacement of
Nisantasi Universitesi, Istanbul, Turkey
scenes or images, movies, and sounds that humans cannot tell apart from real
Correspondence ones. Various technologies have brought the capacity to change a synthetic
Nima Jafari Navimipour, Department of speech, image, or video to our fingers. Furthermore, video and image frauds are
Software Engineering, Haliç University,
Istanbul, 34060, Turkey.
now so convincing that it is hard to distinguish between false and authentic con-
Email: [email protected] tent with the naked eye. It might result in various issues and ranging from deceiv-
ing public opinion to using doctored evidence in a court. For such considerations,
Edited by: Mehmed Kantardzic,
Associate Editor and Witold Pedrycz, it is critical to have technologies that can assist us in discerning reality. This study
Editor-in-Chief gives a complete assessment of the literature on deepfake detection strategies
using DL-based algorithms. We categorize deepfake detection methods in this
work based on their applications, which include video detection, image detection,
audio detection, and hybrid multimedia detection. The objective of this paper is
to give the reader a better knowledge of (1) how deepfakes are generated and
identified, (2) the latest developments and breakthroughs in this realm, (3) weak-
nesses of existing security methods, and (4) areas requiring more investigation
and consideration. The results suggest that the Conventional Neural Networks
(CNN) methodology is the most often employed DL method in publications.
According to research, the majority of the articles are on the subject of video
deepfake detection. The majority of the articles focused on enhancing only one
parameter, with the accuracy parameter receiving the most attention.

This article is categorized under:

Technologies > Machine Learning
Algorithmic Development > Multimedia
Application Areas > Science and Technology

This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any
medium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made.
© 2023 Kadir Has University. WIREs Data Mining and Knowledge Discovery published by Wiley Periodicals LLC.

WIREs Data Mining Knowl Discov. 2024;14:e1520. wires.wiley.com/dmkd 1 of 45

https://doi.org/10.1002/widm.1520
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2 of 45 HEIDARI ET AL.

KEYWORDS
deep learning, deepfake, detection, neural networks, review

1 | INTRODUCTION

The widespread availability of low-cost digital devices, including smartphones, laptops, desktop computers, and digital
cameras, has triggered the development of multimedia material (such as photographs and movies) on the Internet and
wireless communication systems (Feng et al., 2021; Jia et al., 2022; D. Li, Ge, et al., 2020). Furthermore, during the last
few decades, social media has enabled people to quickly communicate recorded multimedia content, resulting in signif-
icant growth in multimedia content output and accessibility (Lv et al., 2020). Increased the speed with which false and
erroneous information can be manufactured and circulated, knowing the truth and trusting the information has
become increasingly challenging, perhaps resulting in disastrous repercussions (Garg & Mago, 2021; A. Li, Spano,
et al., 2020; Niu et al., 2022). So, a deepfake is material created by Deep Learning (DL) (Luo, Yuan, et al., 2022) that
seems genuine in a human's eyes. The term deepfake is a mixture of the phrases DL and fake, and it generally refers to
material created by a deep neural network (DNN), which is a subset of machine learning (ML) (S. Li, Liu, et al., 2020;
Lv, Li, et al., 2021; Y. Wang et al., 2021; Zhong et al., 2021). Everyone can manipulate video and image items (R. Liu
et al., 2021). This has been available for several years due to many user-friendly software packages that allow video
editing, audio, and picture (Aversano et al., 2021). The adoption of smartphone applications that conduct automated
procedures, including lip-syncing, audio instrumentation, and face swaps, has simplified media manipulation
(Tolosana et al., 2020). Furthermore, DL-driven technical developments have resulted in a slew of AI-driven technolo-
gies that make manipulations very convincing and believable (Castillo Camacho & Wang, 2021). Every one of these
techniques seems to be an unquestionably valuable addition to the toolbox of a digital artist (Afchar et al., 2018; Y. Li,
Zhang, et al., 2021). Nevertheless, when used maliciously to create fake media, they may have significant negative social
or personal implications (Muhammad & Hossain, 2022). Deepfakes are a notable example of artificially manipulated
media that has caused serious concern. Furthermore, the deepfake AI-driven technique replaces one person's identity
in a video with that of another (Sanghvi et al., 2021). This often propagates false information by impersonating politi-
cians and distributing revenge porn (Matern et al., 2019).
Deepfake, which refers to various face modification techniques, incorporates cutting-edge technologies, including
computer vision and DL (Lv, Chen, et al., 2021; Masi et al., 2020). Face manipulation is divided into four categories:
expression swap, full-face synthesis, attribute manipulation, and identity swap (Balaji et al., 2021). Identity swap, often
known as face swap, is one of the most common types of deepfake video wherein the faces of the source people are rep-
laced with the faces of the targeted people (Jafar et al., 2020). Deepfake news can also be coupled with forged videos
and photos, making it difficult for users to distinguish them (Matern et al., 2019). This form of deepfake has the poten-
tial to spread over social media and negatively influence people's lives (H. H. Nguyen, Yamagishi, et al., 2019).
Although some deepfakes could be made using classic visual effects or computer graphics, DL methods, including auto-
encoders and generative adversarial networks (GANs), have since been extensively used in the computer vision sector
and are the most recent popular underlying mechanism for deepfake generation (Gosse & Burkell, 2020). Such models
describe a person's emotions and motions and synthesize facial images of someone with similar expressions and actions
(Sanghvi et al., 2021). To train models to generate photo-realistic videos and pictures, deepfake techniques often take
up a considerable amount of audio/video data (Jung et al., 2020). Celebrities and politicians are the first subjects of
deepfakes since they have a vast quantity of videos and photographs available online (Jiang et al., 2021). Deepfakes are
also utilized in pornographic photographs and movies to replace the faces of celebrities and politicians with bodies
(Schlett et al., 2021). Whenever deepfake techniques could be used to make movies of international leaders with fake
statements for falsification reasons, it represents a risk to global security (Fernandes & Jha, 2020). Deepfakes can thus
be used to incite political or religious tensions between countries, deceive the public and influence election results, or
create havoc in financial markets by spreading false information (Jeyaraj & Dwivedi, 2020; Nirkin et al., 2021). How-
ever, it could be used to make artificial satellite photos of the Earth that include features that do not correspond to real-
ity to deceive military analysts, including a fake bridge across a river when there is none in actuality (Biswas
et al., 2021; Heidari, Jafari Navimipour, et al., 2022). It could lead to a troop being misled when crossing a bridge during
combat (Baygin et al., 2022; Siegel et al., 2021). Also, transfer learning (TL) is critical since the model must be trained
in a reasonable period and with little resources to achieve the necessary accuracy for classification over a wide range of
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
HEIDARI ET AL. 3 of 45

instances in the dataset (Garg & Mago, 2021; Kim & Cho, 2021; Zong & Wan, 2022; Zong & Wang, 2022). It enables
models to be trained fast and accurately by finding relatively useful spatial features from large datasets across various
domains at the start of the training process (Heidari, Navimipour, et al., 2022; Luo, Zhang, et al., 2022). As a result, TL
might be a means to combine the necessary computing power and encourage more efficient DL techniques, which
could help with the resolution of several problems (Dwivedi et al., 2021; P. Yu et al., 2021). Because of the size of
datasets, the importance of datasets is significant. The size of the dataset used to generate the findings has a massive
effect. Various datasets, including faceforensics, CelebA, TIMIT, and HFM, are used in this context (Hinton et al., 2015;
Zi et al., 2020).
Nonetheless, there is no thorough and detailed evaluation of the application of DL approaches in the deepfake
detection area that considers all categories such as image, video, and multimedia content. The study focused on empha-
sizing the achievements and detection of deepfakes utilizing DL approaches to combat fraud, stigmas, scams, spoofing,
and so on to offer a complete overview of the modern systems using these technologies. Given recent reviews on the
various DL applications developed for deepfake detection techniques, this review substantially contributes by
addressing the most intriguing study area. This study employs a systematic literature review (SLR) (Doewes et al., 2023;
Zadeh et al., 2023) to find, analyze, and combine findings from related studies. Image deepfake detection, video
deepfake detection, sound deepfake detection, and hybrid multimedia deepfake detection are the four primary kinds of
DL strategies used in deepfake detection. We evaluated numerous features such as advantages, obstacles, dataset,
usages, simulation environments, security, and TL for each category and approach that applied DL methods in this
area. This article discusses the use of DL approaches in deepfake detection and addresses a variety of concerns. We have
also gone over future work in great detail, highlighting all issues to rectify. In a nutshell, the following are the contribu-
tions of this article:

• Providing a wide overview of the topic of DL in the deepfake detection area;

• Providing a full overview of existing methods for DL-deepfake detection;
• Providing an overview of the key methods in DL that incorporate deepfake detection;
• Investigating each strategy for DL-deepfake detection with critical features;
• Highlighting the essential areas where the mentioned techniques can be improved.

The following classification provides the structure of this article. The basic concepts and corresponding terminolo-
gies of DL in deepfake detection are discussed in the next section. The relevant research studies are examined in
Section 3. The research methodology and techniques for article selection are presented in Section 4. The categories of
the selected articles are described in Section 5. The results and comparisons are presented in Section 6. Finally, the open
issues are discussed in Section 7, and the conclusion is defined in the last sections. Moreover, the used abbreviations
are shown in Table 1.

2 | B A S I C CO N C E P T S AN D C O R R E S P O N D I N G T E R M I N O L O G I E S

This section covers the fundamentals of deepfakes, ML, and DL in general and how they are used in the deepfakes
domain.

2.1 | Fake content creation

There are several techniques for manipulating visual material and tracking different targets (Niknejad et al., 2020; Yan
et al., 2022). The following section will provide a quick overview of some of the more common and promising among
them. Adding, replicating, and deleting objects are all very frequent actions (Meskys et al., 2020). A new item can be
added by copying it from another picture (splicing) or the same image (copy-move). An existing item can be removed
by expanding the backdrop to cover it (inpainting), as in the renowned exemplar-based inpainting (S. R. A. Ahmed &
Sonuç, 2021). All of these procedures are simple to carry out with common picture editing software. Furthermore,
appropriate post-processing, including scaling, rotation, or color correction, could be used to match the item to the
scene better, improve the visual appeal, and ensure consistent perspective and scale (Deshmukh & Wankhade, 2021).
Meanwhile, sophisticated computer graphics techniques, DL, and multi-layer networks have produced similar
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
4 of 45 HEIDARI ET AL.

TABLE 1 Table of abbreviations.

Abbreviation Definition Abbreviation Definition

AI Artificial intelligence MFNN Multi-layer fusion neural network
AMTEN Adaptive manipulation traces extraction network ML Machine learning
AUC Area under the curve NLP Natural language processing
AUROC Area under the receiver operating characteristic NN Neural network
BP-DANN Backpropagation based on domain- PoA Proof of authenticity
adversarial neural network
CNN Conventional neural networks RGB Red green blue
DL Deep learning RCNN Region conventional neural networks
DNN Deep neural network RNN Recurrent neural networks
ELA Error level analysis SA-DTH-Net Speaker authentication dynamic
talking habit network-based
GAN Generative adversarial networks SCNN Set CNN
HF Hyperledger fabric SFFN Shallow-FakeFaceNet
HFF Hybrid fake face SLR Systematic literature review
HFM Handcrafted face manipulation SRM Spatial rich model
HMM Hidden Markov model SVM Support vector machine
HRNet High-resolution network TL Transfer learning
JPEG Joint photographic experts group TTS Text-to-speech
LBP Local binary patterns VSA Visual speaker authentication
MCNet Manipulation classification network YOLO You only look once

outcomes with improved semantic consistency in current history (Hossain et al., 2022; Zheng, Liu, & Yin, 2021). Manip-
ulations that do not necessitate the use of advanced AI technologies are frequently referred to as “shallow fakes” or
“cheap fakes” (Dixit & Silakari, 2021). However, its ability to alter perception could be significant (Heidari, Toumaj,
et al., 2022). For instance, a video's meaning can be altered by deleting, adding, or cloning entire sets of frames
(L. Verdoliva, 2020). Aside from these “conventional” manipulations of particular portions of an image or video, DL
and computer graphics are now providing many new ones (Iqbal & Qureshi, 2022). Occasionally media assets are cre-
ated entirely from scratch. To that aim, auto-encoders and GAN enabled effective solutions, particularly for face synthe-
sis, and with a high degree of photo-realism attained (Habeeba et al., 2021). A segmentation map may also be used to
produce a synthetic video or picture. Image synthesis can also be accomplished using simply a drawing or a text
description (Soares & Parreiras, 2020). Similarly, a person's face can be modified depending on an auditory input
sequence (Nasar et al., 2020). Often, the modification alters existing pictures or videos. Style transfer is very well-
conceived since it allows you to alter the style of a painting, convert peaches to oranges, or recreate an image in a new
season. Face manipulation has received much attention because of its high semantic value and wide range of uses
(Kietzmann et al., 2020).

2.2 | Deepfake detection, fake image detection, and fake video detection

Deepfake detection is typically considered a binary arrangement problem in which classifiers are used to distinguish
between reliable and interfering movies (Singh et al., 2021). This technique requires a big library of actual and false
videos to train classification models. The number of fraudulent videos is growing, even though it is insufficient to pro-
vide a standard for validating various finding methods (Q. Liu & Celebi, 2021). Several prior studies on detecting
deepfake images or videos were focused on DL techniques, which could be split into two detection methods: Conven-
tional Neural Networks (CNN)-based methods and Region Conventional Neural Networks (RCNN)-based methods
(X. Zhou et al., 2020). CNN-based techniques take facial pictures from video frames and feed them into the CNN for
training and prediction to get an image-level result. So, in deepfake videos, such algorithms only employ spatial
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
HEIDARI ET AL. 5 of 45

information from a single frame. RCNN-based techniques, on the other hand, require a series of video frames for train-
ing to produce a video-level result. This approach, known as RCNN, combines CNN and Recurrent Neural Networks
(RNN) (Tariq et al., 2021). As a result, RCNN-based techniques could fully use spatial and temporal information in
deepfake movies (Chung et al., 2015). Furthermore, several deepfake detection methods are based on the standard ML
methods, including utilizing a support vector machine (SVM) as a classifier and extracting handmade characteristics,
including biological signals, as a classifier (Yazdinejad et al., 2020). Face swap offers several compelling uses in video
compositing, portrait alteration, and, most notably, individuality protection, as it can exchange faces in photos with
ones from a group of standard images. Nonetheless, it is one of the methods used by imposters to acquire access to
authentication systems. Besides, face photographs have become increasingly hard for forensics as DL, including CNN,
GAN, SVM, random forest, and multi-layer perceptron, can preserve the photos' position, facial appearance, and light.
Among the DL images, those created by GAN are perhaps the most difficult to spot because they are real and high-
quality, thanks to GAN's capacity to learn the supply of input data and produce fresh results with the same input distri-
bution. Though this advancement of GAN continues and various new modifications of GAN are frequently introduced,
most works on the recognition of GAN creating samples do not examine the capability of the recognition models. A step
of image preparation was also used in various methods (Sanghvi et al., 2021). This increases the statistical comparison
between the actual image and the false image at the pixel level, requiring the forensic classifier to examine important
associated and meaningful features with a higher simplification competence than previous image forensics methods or
image stage assessment systems. Also, because of the robust deprivation of frame data following audio-visual compres-
sion, maximum image-based deepfake recognition techniques cannot be applied to videos (Du et al., 2020). Further-
more, videos have a variety of chronological properties that vary between sets of frames, making it challenging for
techniques designed to detect individual fraudulent images. This subcategory focuses on deepfake video recognition
methods and divides them into two groups: those that use chronological characteristics and those that investigate visual
artifacts inside frames (Mehta et al., 2021).
In the meantime, certain deepfake video generation models cannot synthesize all of the textures of the face,
resulting in some somewhat rough fake faces. For example, the small wrinkles on the face cannot be created accurately.
At the same time, the created face is fused into the backdrop during the final phase of creating the false frame. Smooth-
ing techniques are frequently employed to alleviate the border inconsistencies created by this process, which results in
the loss of face texture features. Figure 1a shows a frame from the genuine videos VidTIMIT dataset, whereas Figure 1b
shows a frame from the face manipulation videos DeepFake-TIMIT dataset. Human eyes, following the figure, have dif-
ficulty distinguishing which frame is real. Nevertheless, the actual frame has a more realistic texture, for example, dou-
ble eyelids and wrinkle features around the eyes, but the false frame lacks texture details.

2.3 | Attribute manipulation, expression, and identity swap

The face of one individual in a film is replaced with another person's face in identify swap methods. Two approaches
are often considered: the first is traditional computer graphics-based technologies, such as FaceSwap, and the second is
innovative deepfakes techniques, such as the current ZAO smartphone app (Hongmeng et al., 2020). Also, on Youtube
and commercial websites, you can find quite realistic examples of this type of modification. Such alteration could assist
various industries, including the entertainment industry (Rana & Sung, 2020). On the other hand, it could be used for

FIGURE 1 Real and fabricated frames are shown. (a) is a genuine frame, whereas (b) is a fake (H. Zhao, Zhou et al., 2021).
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
6 of 45 HEIDARI ET AL.

nefarious objectives, including producing celebrity pornographic videos, frauds, and financial fraud, to name a few
(Pokroy & Egorov, 2021). In addition, attribute manipulation, sometimes referred to as face alteration or face ret-
ouching, entails changing aspects of the face, including hair or skin color, gender, age, and the addition of glasses,
among other things (Fink et al., 2020). Then, GAN, including the StarGAN technique introduced in, is commonly used
to carry out this modification procedure. The popular FaceApp smartphone application is an example of this kind of
manipulation (Akhtar et al., 2020). Users might utilize this innovation to test a wide range of products in a virtual envi-
ronment, including cosmetics and makeup, spectacles, and hairstyles (Y. Zhang, Gao, et al., 2021). Also, face reenact-
ment is a type of manipulation that involves altering a person's facial expression. Although several modification
approaches are proposed in the literature, such as at the image level using famous GAN architectures, this group
focuses on the most popular techniques, Face2Face, and NeuralTextures, which substitute one person's facial expres-
sion in a video with that of another person. This form of deception could have catastrophic ramifications if it is
employed (Tu et al., 2021).

2.4 | Creating a fake speech and fake speech detection

Synthetic speech production is a subject that has been explored for many years and has been approached in various
ways (Bekci et al., 2020). As a result, the literature has a vast number of approaches that provide amazing results, and
there is no single consistent method of producing a synthetic voice track. Text-to-Speech (TTS) generation was tradi-
tionally based on concatenative waveform synthesis, which means that provided a sentence as input, the output speech
is created by picking the proper diphone units from a huge dataset of diphone waveforms and integrating them to
ensure acceptability (Katarya & Lal, 2020). Additional post-processing tools help for smoother transitions between
diphones, simulation of human prosody, and a high degree of naturalness (Tjon et al., 2021). The primary disadvantage
of concatenative synthesis is the complexity of changing the tonal features of the sound, for example, to alter the
speaker or embed emotional resonance in the voice. Also, some approaches, such as the Hidden Markov Models
(HMM) based speech synthesis system, are proposed to improve the range of voice characteristics or talking patterns
(Zi et al., 2020). These use contextual HMMs trained on huge datasets of auditory characteristics collected from
diphones and triphones. Another method, called parametric TTS synthesis techniques, seeks to increase the number of
produced voices. In this scenario, a speech signal is produced as an auto-regressive process, given a collection of speech
characteristics, including frequency (Hong et al., 2021), spectral envelope, and excitation signal. Nevertheless, paramet-
ric TTS synthesis yields fewer natural-sounding outcomes than concatenative TTS synthesis (Chi et al., 2020).
Neural networks (NNs) have created natural and versatile synthetic voices (Z. Wang, Ramamoorthy, et al., 2022;
Z. Wang, Ramamoorthy, et al., 2022). Modeling audio sample by sample, for example, has traditionally been thought to
be extremely difficult because speech signals often count 100 of samples per second and preserve essential features at
many timeframes (Nasar et al., 2020). However, in recent decades, CNN and RNN have permitted the construction of
entirely auto-regressive models, allowing the direct generation of raw audio waveforms (Burroughs et al., 2020). It is far
from simple to determine if a speech recording belongs to a genuine person or was created artificially. Nevertheless,
artificial speeches may be created using several methods, each with its own set of characteristics. It is difficult to estab-
lish a comprehensive forensic framework that shows all potential synthetic speech approaches (D. Xie et al., 2020). Fur-
thermore, as ML solutions become more popular, new and improved methods of creating synthetic voice recordings are
regularly presented (A. Li, Masouros, et al., 2021). Keeping up with the growth of speech synthesis literature is therefore
difficult. Even so, the sector has developed a set of detectors to prevent the proliferation of false audio recordings. One
of several aims of this study is to examine deepfake detection in the area of fake speech (Mitra et al., 2020).

2.5 | Domain adaptation and TL

Understanding TL in general and its use in deepfake detection is an important subject that is an intrinsic component of
this area. However, TL is a subfield of DL that leverages information from source domains to help the method acquire
knowledge from the target domain quicker and more effectively. TL has received much interest in the realm of foren-
sics. Loading ImageNet's pre-trained weight to the model before training is a straightforward TL (Marei et al., 2021).
The majority of works utilizing deep domain adaptation rely on discrepancy measurement. In addition, several efforts
are based on domain-adversarial learning and discrepancy assessment (I. Ahmed et al., 2021). So, TL is an approach for
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
HEIDARI ET AL. 7 of 45

solving a new task (apparently related or unrelated) using previously learned model information with little retraining
or fine-tuning. In comparison to traditional ML methods, DL requires a considerable amount of training data. The time
it takes to train for a domain-specific problem is greatly reduced when using TL (Lewis et al., 2020). The dataset is alle-
viated using a TL-based DL approach (Hung & Chang, 2021). Because the growth of deepfakes has been so terrifying
on the internet, detecting them has become a top priority in the current circumstances (Heidari et al., 2022).
Researchers in this sector are working hard to develop some potential TL mechanisms that can assist in alleviating the
challenges in such a circumstance.

3 | RELEVANT REVIEWS

Multiple methods for manipulating material such as photos, video, audio, and so on have been successfully created and
are now accessible to the public, such as FaceSwap, Face2Face, deepfake, and so forth. It allows anybody to quickly
and simply modify faces in video sequences, resulting in amazingly realistic outcomes with no effort. Furthermore, in
this era of fake multimedia, easy access to large-scale public databases and rapid advances in DL techniques, notably
GANs, have resulted in incredibly realistic fake content with societal ramifications. Similarly, deepfake detection is a
significant application of DL and ML that assists detect forgeries in media, including videos and photos. An amount of
research has indeed been done on it, including a thorough study and implementation of many common algorithms
(Albahar & Almalki, 2019; T. T. Nguyen, Nguyen, et al., 2019; Pashine et al., 2021; Saif & Tehseen, 2022; Shelke &
Kasana, 2021; Swathi & Sk, 2021; Weerawardana & Fernando, 2021). So, DL has been recognized as an excellent
method for detecting synthetic media in the deepfake realm. For this reason, L. Verdoliva (2020) examined approaches
for visual media integrity verification or identifying altered pictures and videos. She focused on the rising topic of
deepfakes, or fake media generated by DL algorithms, as well as current data-driven forensic approaches for combating
them. The article examined integrity verification methods, beginning with traditional ways and progressing to DL-based
approaches before concluding with particular deepfake detection algorithms. The review emphasized the limitations of
present forensic techniques, the most pressing concerns, impending difficulties, and future research paths. The work
does not define a method category and does not compare parameters between methods.
Also, Tolosana et al. (2020) presented a detailed overview of facial picture manipulation techniques, particularly
deepfake strategies, as well as methods to identify such alterations. Four methods of facial manipulation are discussed
in detail: (i) complete face synthesis; (ii) identity swap; (iii) attribute manipulation; and (iv) expression swap. They pre-
sent details about manipulation techniques, existing public databases, and essential standards for performance review
of false detection systems, as well as a summary of the results of those evaluations for each manipulation group. They
focused on the newest versions of deepfakes, stressing its advances and problems for fake identification, among all the
topics mentioned in the study. They also talk about unresolved difficulties and future trends that should be explored as
the field progresses.
Mirsky and Lee (2021) specialized in human reenactment and replacement deepfakes. They went into how these
techniques function, the variations in their structures, and what is being done to identify them in detail. They look into
the production and identification of deepfakes and give a detailed look into how these systems work. Their goal is to
give the reader a better grasp of how deepfakes are made and recognized and the latest trends and developments in this
field. They also point out flaws in present defense measures as well as topics that need additional study and emphasis.
Besides, Castillo Camacho and Wang (2021) provided a broad grasp of the detecting methods used in the field of
image forensics. They gathered and presented a variety of DL-based methods grouped into three main categories,
emphasizing the unique characteristics of picture forensics approaches. They discovered that a preprocessing procedure
to achieve a specific feature or a customized initialization on the network's first layer was utilized in many pioneer
works and is still employed in current ones. They provided a detailed overview of image forensics approaches, with a
particular emphasis on profound algorithms. They addressed a wide range of image forensics issues, such as detecting
regular picture alterations, detecting deliberate image falsifications, identifying cameras, classifying computer graphics
images, and detecting developing deepfake images.
Finally, P. Yu et al. (2021) presented the deepfake video production technologies, analyzed the available detection
system, and discussed the path of the research directions. Their paper provided a state of research in deepfake video
detection, namely the production process, various detection algorithms, and current standards. The many sorts of detec-
tion techniques are then discussed. The review relies on new difficulties with current detection algorithms as well as
potential developments. Their study places a strong emphasis on generalization and resilience.
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
8 of 45 HEIDARI ET AL.

Rana et al. (2022) performed a broad analysis in their paper to offer an overview of the research efforts in Deepfake
detection, synthesizing 112 published documents from 2018 to 2020 that provided a range of techniques. They classified
them into four categories: DL-based approaches, traditional ML-based techniques, statistical techniques, and
blockchain-based methods. They also compared the detection capabilities of the various algorithms across different
datasets and concluded that DL-based methods outperform other methods in Deepfake detection.
When these publications were evaluated, it was determined that none of them employed the SLR strategy, and none
of the deepfakes analysis articles had all deepfakes classifications for examining DL methods. Also, according to the
findings, most mentioned publications either focused on a certain component of deepfake administration or described a
specific type of data collection. Furthermore, most of these surveys included no comparison examination and only
looked at less than 20 articles. Our paper focuses on peer-reviewed publications that propose DL algorithms for
deepfake detection tasks, including image, video, sound, and hybrid detection. As a result, we use a slightly different
method than most articles. Table 2 includes a summary of related works. Also, the methodology of the research will be
described in the following section.

4 | M E T H O D O L O G Y O F RE S E A R C H

In the previous section, we discussed some related works that looked into deepfake detection techniques. This
section uses the SLR approach (Esmailiyan et al., 2021; Vahdat & Shahidi, 2020) to comprehend deepfake detection
better. The SLR is a comprehensive examination of all studies on a particular topic. This section provides a detailed
look at how DL approaches are employed in deepfake detection. The validity of the study selection procedures is
next examined. The following subsections describe the search procedure, including research questions and selec-
tion criteria.

4.1 | Question formalization

The study's main goals are to identify, distinguish, assess, and evaluate all relevant publications found in deepfake
detection methods from DL. An SLR can be used to explore the aspects and features of the methodologies to attain the
aims specified before. Another goal of SLR is to understand better the major concerns and problems that this industry
encounters. A few Research Questions (RQs) that have been defined are as follows:

RQ 1. What are the applications of DL in deepfake detection?

This question was addressed in Section 1.

RQ 2. What are DL methods in deepfake, and what usages do they have?

This question was answered in Section 2.

RQ 3. Is there any study that has been published as a review article in this field? What distinguishes this
article from past research?

This question was answered in Section 3.

RQ 4. What are this area's leading issues and unanswered problems?

Section 5 will present the answers to this topic, while Section 7 will present the open problems.

RQ 5. How can we find the article and choose the deepfake detection DL methods?

This is dealt with in Section 4.2.

 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
HEIDARI ET AL. 9 of 45

TABLE 2 Compilation of related works.

Researchers Contributions Scope Advantage Weakness

L. Verdoliva (2020) Presenting an overview of Fake media • Deepfake's backstory is • There has not been
contemporary presented. any in-depth review
manipulation • Issues and possible of the articles.
techniques solutions are explored.
Tolosana et al. (2020) Examining face-altering Image deepfake • Different criteria for • It is unclear how
techniques detection evaluating articles are articles are chosen for
taken into account. review.
Mirsky and Lee (2021) Providing deepfake Deepfake in general • Challenges and • It is unclear how
creation and detection potential guidance are articles are chosen for
services discussed. review.
Castillo Camacho Examining DL-based Image forensic • Taking into account all • It is unclear how
and Wang (2021) image forensic methods aspects of the criteria articles are chosen for
for image forensics. review.
P. Yu et al. (2021) Focusing on deepfake Deepfake video • An in-depth description • There is no
video detection, its of future work. comparison between
history, current • In-depth examination the articles.
research, and plans. of datasets.
(Rana et al., 2022) Demonstrating several DL-ML and statistical • There is a comparison • There is no
cutting-edge deepfake models between the articles. discussion of all kinds
algorithms. of deepfake
applications.
Ours Providing a DL-ML methods in the • An in-depth description • Papers published
comprehensive review video, image, audio, of future work. before 2018 are not
of the literature on and hybrid • In-depth examination allowed.
deepfake detection multimedia of datasets.
techniques based on detection • Challenges and
DL-based algorithms potential guidance are
discussed.

RQ 6. How can the DL mechanisms in multimedia deepfake detection be classified? What are some of their
examples?

The answer to this question can be found in Section 5.

RQ 7. What procedures do the researchers employ to conduct their research?

This question is answered in Sections 5.1–5.4.

4.2 | The process of article selection

This study's article search and selection process is divided into four sections. This is depicted in Figure 2. Table 3 shows
the keywords and terms used to search the publications in the first step. The articles in this collection are the result of
an electronic database search. Some of the used electronic databases are Google Scholar, Scopus, IEEE, ACM, Springer,
Elsevier, Emerald Insight, Taylor & Francis, Wiley, Peerj, and MDPI. The other items uncovered are journals, confer-
ence articles, books, chapters, notes, technical studies, and special issues. The first stage delivered 651 articles. In addi-
tion, the distribution of articles by the publisher is depicted in Figure 3.
The ultimate number of articles to study is determined in Stage 2 through two processes. The articles are initially
evaluated (stage 2.1) using the criteria outlined in Figure 4. There are currently 254 articles left. In Figure 5, the pub-
lisher's distribution of articles is illustrated. Besides, stage 2.2 excludes review articles; at the previous stage, 8.87% of
the remaining 254 articles were review articles. IEEE publishes the majority of research publications (21%). Springer
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
10 of 45 HEIDARI ET AL.

and Elsevier publish the majority of review articles (5.54%). There are currently 210 articles available. In Stage 3, the
titles and abstracts of the articles were reviewed. The articles' methodology, assessment, discussion, and conclusion
have all been double-checked to ensure that they are relevant to the study. At this time, 88 articles have been selected
for further review. Finally, 34 articles were picked to review and examine the other publications since they matched the
tight criteria. The distribution of the chosen articles by their publishers is depicted in Figure 6. IEEE publishes
the majority of the selected articles. Wiley, IGI Global Publishing, and Tech Science Press have the lowest. The number
of publications published in 2021 and 2020 is equal (50%, 16 articles). The journals that publish the publications are
shown in Figure 7. The IEEE Access journal publishes the most articles (9.4%, three articles). Table 4 also lists the

FIGURE 2 The steps involved in the article search and selection procedure.

TABLE 3 Keywords and search phrases.

S# Search terms and keywords

S1 “Deep learning” and “Deepfake detection”
S2 “Machine learning” and “Deepfake detection”
S3 “Deep learning” and “Image forensic”
S4 “Neural network” and “Deepfake detection”
S5 “AI methods deepfake detection” or “Artificial intelligence deepfake”
S6 “Deep transfer learning” and “Deepfake”
S7 “Transfer learning” and “Deepfake detection”

FIGURE 3 The first stage is the distribution of the articles by the publisher.
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
HEIDARI ET AL. 11 of 45

FIGURE 4 Inclusion criteria for the paper selection process.

FIGURE 5 The articles are provided by the publishers at Stage 2.1.

chosen articles' specifications (updated on July 19, 2021). In addition, in the following part, the deepfake detection
mechanism is explored in-depth, and its properties are discussed.

5 | DEEP FAKE DETECTION MECHANISMS

In Section 4, we explored how we choose articles and the factors that are significant to us, and then we listed the arti-
cles chosen and appraised based on their merits. In this section, the DL approaches for identifying deepfake, and associ-
ated circumstances are discussed in this section. In this part, 32 articles will be discussed, all of which fulfill our
selection criteria. First, we divide the techniques into four categories based on their intended use: Figure 8 displays the
suggested taxonomy of DL-deepfake detection techniques, including image, video, sound, and hybrid multimedia
content.
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
12 of 45 HEIDARI ET AL.

FIGURE 6 The distribution of the selected articles by the publishers.

FIGURE 7 Distribution of the selected articles based on journals.

5.1 | Fake image detection

As described in Section 2, Fake face image detection is the most difficult challenge in the field of image forgery detection.
Phony photos can construct fake identities on social media platforms, allowing for the illicit theft of personal information.
The fake picture generator, for example, can be used to create photos of celebrities with inappropriate content, which
might be dangerous. This section will look at detecting fake images using DL approaches. Deepfake's development signifi-
cantly lowers the bar for face fabrication techniques. Deepfake, in particular, uses GANs to substitute an original image's
face with another person's face. Because the GAN models were trained on 10 of 100 of photos, they are more likely to pro-
duce realistic faces that can be accurately spliced into the primary image. The image may be made more real by using
appropriate post-processing. Furthermore, with the growth of DL, face-swap innovation has been employed in various set-
tings, encompassing privacy protection, multimedia synthesis, and other novel applications (M. Zhang, Chen, et al., 2021;
M. Zhang, Chen, et al., 2020). So, W. Zhang, Zhao, et al. (2020) investigated a comprehensive DL-based counterfeit feature
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
HEIDARI ET AL. 13 of 45

TABLE 4 The chosen articles' specifications (updated on 15 January, 2022).

JCR based Scopus Country of H-index

Publisher Author Year Citation 2021 based 2021 Journal name journal based 2021
MDPI W. Zhang, Zhao, 2020 7 Q2 Q2 Entropy Switzerland 74
and Li (2020)
Elsevier Lee et al. (2021) 2021 10 Q1 Q1 Applied Soft Computing Netherlands 143
Elsevier Guo et al. (2021) 2021 8 Q2 Q1 Computer Vision and USA 138
Image Understanding
IEEE Guarnera et al. 2020 14 Q2 Q1 IEEE Access USA 127
(2020)
IEEE I.-J. Yu et al. 2021 4 Q2 Q1 IEEE Access USA 127
(2020)
Elsevier J. Yang et al. 2021 2 Q1 Q1 Future Generation Netherlands 119
(2021) Computer Systems
MDPI Hsu et al. (2020) 2020 53 Q2 Q2 Applied Sciences Switzerland 52
IEEE (Güera & 2018 608 - - IEEE International USA -
Delp, 2018) Conference on
Advanced Video and
Signal Based
Surveillance
Elsevier X. H. Nguyen 2021 5 Q2 Q1 Forensic Science Ireland 120
et al. (2021) International
IEEE Jung et al. 2020 36 Q2 Q1 IEEE Access USA 127
(2020)
- Karandikar 2020 2 - - International Journal of India 18
et al. (2020) Advanced Trends in
Computer Science and
Engineering
IGI Global Zhao, Wang, 2021 0 - Q4 International Journal of USA 15
Publishing and Lu (2021) Digital Crime and
Forensics
Elsevier Z. Xu et al. 2021 1 Q2 Q1 Journal of Visual USA 81
(2021) Communication and
Image Representation
Springer Kohli and Gupta 2021 0 Q2 Q1 Multimedia Tools and Netherlands 70
(2021) Applications
Hindawi Chen and Tan 2021 0 Q4 Q2 Security and Egypt 43
(2021) Communication
Networks
IEEE A. Yan, Fan, 2021 4 Q1 Q1 IEEE Transactions on USA 168
et al. (2021) Circuits and Systems for
Video Technology
Elsevier Caldelli et al. 2021 3 Q2 Q1 Pattern Recognition Netherlands 57
(2021) Letters
John Wiley L. Yan, Yin-He, 2021 3 Q1 Q1 International Journal of United Kingdom 87
et al. (2021) Intelligent Systems
IEEE Mitra et al. 2020 3 - - International Symposium USA -
(2020) on Smart Electronic
Systems (iSES)
IEEE Suratkar et al. 2020 1 - - International Conference USA -
(2020) on Computing,
Communication and
Networking
Technologies (ICCCNT)
(Continues)
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
14 of 45 HEIDARI ET AL.

TABLE 4 (Continued)

JCR based Scopus Country of H-index

Publisher Author Year Citation 2021 based 2021 Journal name journal based 2021
IEEE Bonettini et al. 2021 82 - - International Conference Italy -
(2021) on Pattern Recognition
IEEE Cozzolino et al. 2019 42 - - IEEE/CVF Conference USA -
(2019)
Springer Borrelli et al. 2021 2 Q3 Q2 Eurasip Journal on USA 88
(2021) Advances in Signal
Processing
IEEE C.-Z. Yang et al. 2020 8 Q1 Q1 IEEE Transactions on USA 133
(2020) Information Forensics
and Security
ACM Mittal et al. 2020 40 - - Proceedings of the 28th USA -
(2020) ACM international
conference on
multimedia
ACM R. Wang et al. 2020 21 - - in Proceedings of the 28th USA -
(2020) ACM International
Conference on
Multimedia
IEEE Wijethunga 2020 2 - - International Conference USA -
et al. (2020) on Advancements in
Computing (ICAC)
MDPI Khalil et al. 2021 3 - Q2 Future Internet Switzerland 28
(2021)
IEEE Chintha et al. 2020 28 Q1 Q1 IEEE Journal of Selected USA 120
(2020) Topics in Signal
Processing
IEEE Kong et al. 2021 3 Q1 Q1 IEEE Transactions on USA 168
(2021) Circuits and Systems for
Video Technology
IEEE Sun et al. (2020) 2020 1 - - IEEE International USA -
Workshop on
Information Forensics
and Security
IEEE Nasar et al. 2020 0 - - Recent Advances in USA -
(2020) Intelligent
Computational Systems
(RAICS)
ACM Chugh et al. 2020 16 - - Proceedings of the 28th USA -
(2020) ACM International
Conference on
Multimedia
IEEE Chan et al. 2020 3 - - IEEE/ITU International USA -
(2020) Conference on Artificial
Intelligence for Good
(AI4G)

extraction approach that could successfully differentiate between real and fake pictures generated using DL. Their forgery
feature extraction approach may disclose face-swap pictures based on DL and error level analysis (ELA), which can differ-
entiate DL-generated facial images successfully. Instead of the original photos, they utilized these ELA images for training
the CNN model. Converting the original picture to an ELA image is one way to increase the CNN model's training effi-
ciency. Because the ELA picture has less information than the original image, this could enhance efficiency. The ELA
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
HEIDARI ET AL. 15 of 45

FIGURE 8 The suggested DL-deepfake detection taxonomy separated four distinct methods.

image feature concentrates on the part of the original picture in which the error level exceeds the threshold value. Fur-
thermore, the pixels in the ELA picture are frequently significantly distinct from the neighboring pixels, and the contrast
is quite visible; therefore, the image processed by ELA develops the effectiveness of the training CNN model. They trained
a CNN model to extract the counterfeit aspects of the ELA photos and then determined whether or not the input image
was fake. They examined the approach and demonstrated its usefulness in practice. Because the ELA images created in
the previous phases may emphasize the features of the original image when the error level is greater than the threshold
value, just two convolution layers are necessary. As a result, it is simpler to isolate counterfeit characteristics and evaluate
if the image is authentic. It demonstrates that the characteristics in the picture analyzed by ELA could be effectively
employed to identify image authenticity. Additionally, their tests showed that the ELA technique greatly improves the
training efficiency of the CNN model. They assessed the strategy and demonstrated its efficacy in practice. The findings
showed that the characteristics in the picture processed by ELA might effectively be utilized to determine image authen-
ticity. Additionally, their research shows that the ELA approach may considerably improve the training efficiency of the
CNN model. Especially, the amount of processing may be greatly decreased without any loss of accuracy if the needed
floating-point computing power is lowered by more than 90%.
Also, handcrafted face manipulation (HFM) images dataset and soft computing neural network models
(Shallow-FakeFaceNets) with an effective facial manipulating detection process were presented by Lee et al. (2021).
Shallow-FakeFaceNet (SFFN), the neural network classifier model, demonstrates the capacity to recognize false pic-
tures by focusing on altered facial landmarks. The detection process simply uses Red Green Blue (RGB) information to
detect fraudulent facial pictures and does not use any metadata, which could be readily altered. The technique performs
better in Area Under the Receiver Operating Characteristic (AUROC), gaining 3.99% F1-score and 2.91% AUROC on
identifying handcrafted fake facial images, and 93.99% on detecting small GAN-generated fake images, gaining 1.98%
F1-score and 10.44% AUROC. The approach is intended to develop an automated defensive system to battle false pic-
tures used during various online services and apps, utilizing their cutting-edge HFM and SFFN. Furthermore, the study
offers a variety of test findings that could be used to successfully drive future applied soft computing studies to battle
and successfully identify human and GAN-generated false face pictures.
Besides, as preprocessing, Guo et al. (2021) developed a simple yet successful Adaptive Manipulation Traces Extrac-
tion Network (AMTEN) component that uses the convolution layer as a predictor to retrieve photo manipulation traces.
During the backpropagation pass, the weights are dynamically adjusted. Traces were repeated in successive layers to
optimize manipulation traces. They additionally, created a false face detector, AMTENnet, by combining AMTEN with
CNN. AMTEN's manipulation traces are put through CNN to develop different discriminative characteristics. A set of
tests were carried out in which many common post-processing processes were chosen to replicate actual forensics in
complicated circumstances. Their findings demonstrate that AMTENnet obtained higher detection accuracy as well as
desired generalization capabilities. Also, AMTENnet outperformed other approaches in terms of average detection
accuracy on the Hybrid Fake Face (HFF) dataset by 7.61%. Indeed, AMTEN is the primary advantage here, as it pro-
duces superior residual extraction than Constrainedconv and Spatial Rich Model (SRM). It is indeed worth noting that
AMTEN can be used as a simple residual predictor for additional face forensic procedures.
Also, Guarnera et al. (2020) presented the completion of a previous study on deepfake image analysis. An
expectation–maximization technique is used to extract the Convolutional Traces (CTs): a unique fingerprint that can be
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
16 of 45 HEIDARI ET AL.

used to determine if a photo is a deepfake and the GAN architecture that produced it. The retrieved CT is a fingerprint
with strong discriminative power and resistance to attacks and is independent of high-level picture ideas. The results
also revealed that a simple and fast-to-compute method could outperform the state-of-the-art. Furthermore, CT is con-
nected to the image production process, and better performance could well be achieved by rotating input pictures to
locate the most significant direction. With an overall classification accuracy of more than 98% and deepfakes from
10 distinct GAN architectures included in pictures of faces, the CT proves to be trustworthy and independent of image
semantics. Lastly, testing on deepfakes created by the FACE application, which achieved 93% accuracy in the deepfake
detection task, proved the efficiency of the method in a real-world setting.
Besides, I.-J. Yu et al. (2020) introduced the Manipulation Classification Network (MCNet) to categorize the various
manipulation techniques used on Joint Photographic Experts Group (JPEG) compressed pictures by utilizing several
domain characteristics. To optimize the efficiency of manipulation classification for JPEG compressed images, the
MCNet uses spatial, frequency, and compression domain data. They created domain-specific preprocessing and network
topologies. The results indicate that the MCNet outperforms existing manipulation detection networks as well as
numerous low-level vision networks. The suggested network is excellent for fine-tuning current forensics tasks such
as deepfake detection and JPEG picture integrity verification since it considers many alternative alteration methods.
The results specify that the MCNet outperforms existing manipulation detection networks as well as numerous low-
level vision networks. The suggested network is excellent for fine-tuning current forensics tasks such as deepfake detec-
tion and JPEG picture integrity verification since it considers many alternative alteration methods.
J. Yang et al. (2021) recognized the tiny textural variations between the actual and fake facial images and used the
picture saliency approach to demonstrate them. Depending on this finding, they use the enhanced guided filter to do
picture preprocessing on all false and real photos, intending to improve the texture artifacts included in the face modifi-
cation image, which we term guided characteristics. However, this distortion cannot be caught visually; these magnified
texture variations will be learned using a Resnet18 network that can constantly downsample and resample to identify
actual and false facial pictures accurately. Studies showed that their technique has the best detection accuracy in both
whole picture and face image training, which is at the cutting edge of current research. Simultaneously, due to the
extremely expanded texture characteristics, the network could record differences rapidly and correctly, achieving con-
vergence quicker, and reducing training time. Nevertheless, their approach has certain flaws, which is a frequent issue
in investigating real and fake face recognition. Network model training still needs enormous amounts of data to obtain
reasonable accuracy. Rather than building a universal detection network, training a new authentication network is still
required for the unknown face tampering approach.
Finally, Hsu et al. (2020) proposed a fake feature network-based paired learning method to recognize fake faces and
generic pictures created by state-of-the-art GANs. They can learn middle- and high-level and discriminative false fea-
tures by combining cross-layer feature representations. The reduced DenseNet is then transformed into a two-streamed
network structure that can accept paired data as input. The suggested network is then trained to employ paired learning
to differentiate between the characteristics of the fake and real photos. Lastly, a classification layer is added to the
suggested common fake feature network to determine if the input picture is real or false. Their paired learning system
permits fake feature learning, allowing the trained fake image detector to recognize the false picture created by a new
GAN even though it was not included in the training phase. The findings showed that their technique beat other state-
of-the-art methods in terms of precision and recall rate. Table 5 discusses the deepfake image applications used in
deepfake and their properties.

5.2 | Fake video detection

As described in Section 2, one of the most difficult problems is detecting video deepfakes. To create a very advanced
face-swap movie using deepfake, only one GPU and a vast amount of training data are required. Several technology
enthusiasts have posted a large number of live-performance films on the short video platform that replace regular peo-
ple's faces with synthetic stars' faces, resulting in a topic to be debated. Concerns have been raised about this technique.
Before this technique, it was widely assumed that videos were dependable and that they could even be used as video
evidence in multimedia forensics. In the digital age, video deepfake technology has posed a threat to public confidence.
Some even believe that this innovation would stifle societal growth. Thanks to the rapid rise in the availability of open-
source datasets, significant advancements in the study of topics like GANs, and significant technological developments
in the field of high-speed computing, creating and manipulating fake videos has become a relatively simple task in
 TABLE 5 Image deepfake techniques' approaches, attributes, and features.

Security Simulation Using

Authors Main idea Advantages Research challenges used? Dataset environment TL? Method Usage?
HEIDARI ET AL.

Zhang, Zhao, Using a binary classifier • The achieved accuracy • Poor robustness No Milborrow Python No CNN Image
and Li (2020) trained by a CNN. was 97% University of
• The AUC stood Cape Town
at 97.6% dataset
Lee et al. (2021) Suggesting an approach • Obtaining 72.52% • Robustness is not No HFM dataset Keras Python No GAN Image
with an effective end- AUROC taken into account DL library
to-end false face • Obtaining 93.99%
detection pipeline that accuracy on difficult
can identify fake face low-resolution pictures
pictures.
Guo et al. (2021) Providing a preprocessing • AMTENnet achieves an • Detector's No HFF dataset Caffe framework No CNN Image
module named AMTEN average accuracy of up robustness is low
for face image forensics. to 98.52%
• Achieves desirable
preprocessing
Guarnera et al. Presenting an • Obtaining a 93% • High delay No The FACE APP, a Keras No Expectation– Image
(2020) expectation– accuracy rate • High energy dataset, and maximization
maximization method • In a real-world consumption CELEBA images. + CNN
trained to identify and scenario, efficacy is
extract a fingerprint. demonstrated
• High robustness
I.-J. Yu et al. Presenting the MCNet to • High robustness • High complexity No ALASKA dataset PyTorch Yes CNN Image
(2020) utilize multi-domain • High accuracy • High energy
spatial, frequency, and consumption
compression domain
characteristics.
J. Yang et al. Using the image saliency • Detection accuracy is • Poor robustness No Faceforensics++ Pytorch No CNN + simple Image
(2021) to determine the texture 0.9990 dataset linear iterative
depth and pixel clustering
difference between
actual and fake facial
images.
Hsu et al. (2020) Presenting an image • Achieving a modest • High energy No CelebA Not mentioned No Pairwise Image
detector comprised of an level of precision and consumption learning
enhanced DenseNet recall • Low robustness
backbone network and
Siamese network
architecture.
17 of 45

19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
18 of 45 HEIDARI ET AL.

terms of the cost of creating a manipulated sequence. By superimposing a politician's face over the face of a target actor,
a deepfake film might be used to slander and promote false political propaganda. Also, it could then be used to embar-
rass, shame, or cause a schism among major political institutions, jeopardizing national peace by disrupting the delicate
balance of peace between people, governments, and nations worldwide. So, detecting such fraudulent videos has
become incredibly important. More academics have begun to look at detecting techniques and the distinctions between
authentic and phony films. Traditional techniques and DL approaches have been the most commonly employed. Addi-
tionally, living body recognition in face recognition and multimedia forensics can provide options. It is thought that DL
can currently build realistic phony faces, detect and investigate invisible forgery traces, and recognize forgery films.
DL approaches, as opposed to standard picture forensic techniques, incorporate feature extraction and feature classifica-
tion into a network structure and achieve an end-to-end effective automatic feature learning classification methodology
(Zheng, Liu, Ni, et al., 2021). On the other hand, typical image forensic approaches are typically unsuitable for video
since compression significantly reduces data quality. As DL improves in digital forensics, it upends standard image
forensic methodologies. As a result, automated approaches for detecting deepfake movies have become extremely essen-
tial, given the global influence and the extent to which these videos could harm peace and security. In this section, we
will look at how to detect fake videos using DL approaches. So, Güera and Delp (2018) presented a temporal-aware
method for detecting deepfake films automatically. They demonstrated a temporal-aware system for detecting deepfake
videos automatically. They demonstrated a robust trainable recurrent deepfake video identification system. For
processing frame sequences, the suggested system employs a convolutional LSTM structure. A convolutional LSTM has
two important components: (1)_ CNN for extracting frame features. (2) Temporal sequence analysis using LSTM. Given
an unknown test sequence, they retrieved a collection of features produced by the CNN for each frame. The characteris-
tics of numerous consecutive frames were then concatenated and sent to the LSTM for analysis. Ultimately, they calcu-
lated the odds of the sequence being a deepfake or a non-manipulated video. These attributes are then used to train an
RNN to determine whether or not a video has been manipulated. They tested their strategy against a vast collection of
deepfake films gathered from various video websites. They demonstrate how, despite a minimal architecture, the sys-
tem can obtain competitive outcomes in this task. Their experiments with a large collection of digitally altered videos
revealed that a simple convolutional LSTM structure could indeed effectively forecast whether a video has been manip-
ulated or not with as little as 2 s of video data.
Also, X. H. Nguyen et al. (2021) proposed creating three-dimensional (3D) pictures as well as a 3D CNN model that
can learn characteristics of spatial and temporal dimensions from the 3D images. They used 3D convolution kernels to
create a deep 3D CNN to learn spatio-temporal characteristics in short consecutive frame sequences to recognize
deepfake videos. In their method, the feature maps in the convolutional layers are linked to the successive frames of
the previous layer. As a result, it can collect information about faces in spatial and temporal dimensions in a series
of frames. Studies showed that their suggested model performs very well on deepfake FaceForensic and VidTIMIT
datasets. The suggested model excelled on both datasets by more than 99% due to the elevated videos.
Besides, Jung et al. (2020) devised a way to detect deepfakes created by the GANs model by analyzing substantial
variations in eye blinking, a spontaneous and unconscious human behavior. Blinking patterns differ depending on a
person's gender, age, cognitive activities, and time of day. Consequently, their system detected these changes utilizing
ML, various methods, and a heuristic way to verify the integrity of deepfakes. Their method, which was developed uti-
lizing prior experiments' findings, constantly showed a substantial possibility of validating the integrity of deepfakes
and regular videos, correctly recognizing deepfakes in seven of eight videos (87.5%). Nevertheless, one study drawback
is that blinking is also associated with mental illness and dopamine activity. The integrity verification may not be rele-
vant to persons suffering from mental disorders or who have issues with nerve conduction pathways. Conversely,
because cyber-security attacks and defenses are always evolving, this may be enhanced by a variety of techniques. The
suggested technique points towards a new way to overcome the limitation of integrity verification algorithms.
However, Karandikar et al. (2020) suggested technique employs TL on a CNN model to train the dataset and focuses
on face modification to identify counterfeit. Their classifier is built on a VGG-16 model, which is then supplemented by
batch normalization, dropout, and a proprietary two-node dense layer. The two nodes in the suggested architecture's
last dense layer are being used for two final classes (real and fake). The batch normalization layer is used to normalize
and scale the inputs from the preceding layer. Dropout is also included to prevent overfitting and improve weight opti-
mization. At each epoch, the dropout layer will arbitrarily transmit some nodes as off from the preceding layer. It will
result in better training because this layer introduces some unpredictability during weight updates. Their method per-
formed well and can effectively gather features necessary for additional processing to test for deepfakes. The suggested
model's accuracy falls with low-quality pictures, and with medium-quality videos, the accuracy must be enhanced
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
HEIDARI ET AL. 19 of 45

further by utilizing mixed models for training on temporal parameters. As a result, better training will result from a bet-
ter dataset of higher quality.
In addition, Z. Zhao, Zhou et al. (2021) presented a method for detecting face fraud in deepfake videos that is more
accurate. This has offered a Multi-layer Fusion Neural Network (MFNN) to catch diverse artifacts caused by deepfake
forgeries in three distinct layers. The technique offered a deepfake detection method by combining feature maps gener-
ated from multiple levels in the detection network. They constructed shortcut connections for feature maps from vari-
ous layers, reduced their size, and sent them directly to the final layer to categorize the extracted features. The features
maps collected by the shallow, medium, and deep layers are output and fused simultaneously during categorization.
The FaceForensics++ dataset is also used to train and test the network. The findings demonstrate that the novel detec-
tion approach beats the previous methods in detecting deepfake videos. MFNN has significantly increased accuracy in
identifying low-quality deepfake movies, which is generally more challenging for previous techniques.
Z. Xu et al. (2021) proposed a perspective on the faces of numerous video frames as a set to examine facial altered
video recognition and a unique framework Set CNN (SCNN). Their system provides three examples: t-MesoNet,
t-XceptionNet, and t-XceptionNet. Also, comprehensive studies show demonstrated the approach outperforms prior
techniques and may be used as a foundation for combining with other backbone networks. Nevertheless, they discov-
ered that a stronger backbone network could produce superior outcomes. As a result, seeking a better backbone net-
work could be the next step in the right way, and as well as the set reduction technique is critical for SCNN. The three
distinct set reduction techniques are straightforward and intuitive in this case.
Also, Kohli and Gupta (2021) presented a strategy for exploiting the frequency domain characteristics of face forg-
ery. A Frequency CNN (FCNN) is used in their technique to evaluate and categorize clean and counterfeit faces. To
assess the efficacy of the FCNN, the FaceForensics++ dataset is employed. Studies showed that FCNN detects forgeries
efficiently in actual circumstances, including such high and low video quality. Furthermore, the display of activation
maps demonstrates that FCNN learns different frequency characteristics for deepfake, Face2Face, and FaceSwap
manipulation methods. Among all other facial modification techniques, the FCNN detects deepfakes with the greatest
recall of 0.9256, 0.8639, and 0.8399 for raw, c23, and c40, respectively. This technique is also tested on a Celeb-DF
(v2) dataset as well as an automated FaceForensic benchmark. The findings demonstrate the usefulness of the
suggested approach for detecting face manipulation.
Likewise, Chen and Tan (2021) introduced feature transfer, a two-stage deepfake detection approach that relies on
unsupervised domain adaption. The feature vectors derived from CNN are utilized in Backpropagation Based
on Domain-Adversarial Neural Network (BP-DANN) for adversarial TL, leading to higher efficiency than end-to-end
adversarial learning. The face detection network is initially utilized in the preprocessing stage to extract the face region
of the video frame, which is then enlarged by 1.2 times to crop and save the face picture. The CNN is trained on the
large-scale deepfake dataset, which a third party provides. Lastly, the facial images are input into the CNN to obtain
the 2048-dimensional feature vectors. The collected feature vectors are stored to easily load them into the BP-DANN for
the training of unsupervised domain adaptive (B. Xie et al., 2023). Furthermore, the features extraction CNN pre-
trained on a big deepfake dataset might well be utilized to extract additional transferrable feature vectors, reducing the
gap between the source and target throughout unsupervised domain adaptive training.
Also, A. Yan, Yin-He, et al. (2021) worked on compressed deepfake movies with a low-quality factor to cater to scenar-
ios commonly seen on social media. In reality, compressed videos are widespread on social media platforms like
Instagram, Wechat, and Tiktok. As a result, determining ways to detect compressed deepfake films becomes a critical
challenge. In addition, they use the frame-level stream with a low complexity network and prune the model to avoid
fitting the noise because of the noise introduced by compression. The temporality-level stream is used to extract the incon-
sistency between frames to discover the temporal characteristics of compressed movies. The two streams extract com-
pressed video's frame-level and temporal-level information. They tested their suggested two-stream technique on
deepfakes, FaceSwap, Face2Face, NeuralTextures, and Celeb-DF datasets, and the results outperformed previous work.
The accuracy of cross-compression detection findings demonstrates that their approach is robust to compression factors.
Furthermore, Caldelli et al. (2021) suggested optical flow field dissimilarities differentiate between deepfakes and
real videos using CNN. Their research is based on the usage of CNNs that have been trained to detect potential motion
dissimilarities in the temporal structure of a video clip using optical flow fields. The test results produced on the
FaceForensics++ dataset are intriguing and demonstrate that this feature is well suited to extract distinctive character-
istics between the fake and real instances, particularly when dealing with the tough cross-forgery scenario. Moreover, it
is demonstrated how this technique leverages discrepancies on the temporal axis, which improves their efficiency when
integrated with well-known state-of-the-art frame-based approaches.
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
20 of 45 HEIDARI ET AL.

L. Yan, Yin-He, et al. (2021) proposed a lightweight 3D CNN. The CT module is intended to extract features with
fewer parameters at a higher level. 3D CNNs are used as a spatial-temporal module to merge spatial information in the
time dimension. However, SRM characteristics are collected from the input frames to suppress frame content and
emphasize frame texture, allowing the spatial-temporal module to function better. Also, the CT module seeks to extract
deep-level features from the outcome of the spatial-temporal module using as few parameters as possible. As previously
stated, 3D CNNs have larger parameters than traditional two-dimensional (2D) CNNs, which harms convergence and
generalization capabilities. The CT module is intended to replace the standard 3D convolution layers to minimize the
number of parameters. The results demonstrate that their network has fewer parameters than other networks and out-
performs existing state-of-the-art deepfake detection techniques on major deepfake data sets. Because their methodol-
ogy overcomes the problem of high deployment consumption while retaining good detection performance, it is clear
that deepfake detection on edge devices will be used soon.
Also, Mitra et al. (2020) demonstrated a DL-based method for detecting deepfake videos on social media with excel-
lent accuracy. They categorize and modify footage using a neural network-based technique. A model is developed that
consists of a CNN and a classifier network. The CNN modules were chosen from three different structures:
XceptionNet, InceptionV3, and Resnet50, and a comparison study was conducted. They analyzed three existing CNN
modules and decided on Xception net as the feature extractor for the most accurate model, along with the proposed
classifier. They used intermediate compression to train the network and achieved good accuracy even in a high-loss
environment. Even without training with extremely compressed videos, their approach is the key to achieving great
accuracy. The amount of frames retrieved from the movie determines the algorithm's complexity.
Suratkar et al. (2020) developed a system for detecting false videos that use a CNN architecture and the TL method-
ology. Their approach uses a CNN to collect features from each movie frame to build a binary classifier that can effi-
ciently distinguish between real and altered videos. The process is tested on many deepfake films culled from diverse
datasets. So, the findings demonstrate that using TL to develop a relatively robust model for deep fake detection is
doable. With the notion of TL at its helm, the technique employed allows any model to be trained significantly faster,
resulting in a significant reduction in training time. TL also makes it easier to transfer valuable knowledge from one
work to another, which was demonstrated in this situation. To obtain generality in mind, their paper aimed for maxi-
mal generality by combining datasets collected using various methodologies from various sources.
Bonettini et al. (2021) handled the difficulty of detecting face alteration in video sequences by means of recent facial
manipulation techniques. They investigated the assembly of several trained CNN models. In the suggested method, sev-
eral models are created by beginning with a base network (EfficientNetB4) and employing two distinct concepts:
(i) attention layers and (ii) siamese training. On two publicly available datasets with over 119,000 videos, they demon-
strated that merging these networks yields promising face modification detection results.
Finally, Cozzolino et al. (2019) developed a neural network that enhanced the model-related traces concealed in a
video, extracting a form of camera fingerprint known as video noiseprint, inspired by recent work on photos. The net is
trained using a Siamese method on pristine films, limiting distances between same-model patches and maximizing dis-
tances between unrelated patches. Experiments demonstrated that systems based on video noiseprints work well in
important forensic tasks, including camera model identification and video forgery localization, with no requirement for
previous information or fine-tuning.
However, Table 6 lists the deepfake video applications that use DL techniques and their features.

5.3 | Fake sound detection

Section 2 describes voice deepfake detection as a mostly unexplored field. Machine-generated voices mostly inhabit the
daily lives. People are more likely to control daily tasks using speech as technology becomes more automated. Virtual
assistants such as Google Assistant, Alexa, Siri, Bixby, and others increasingly use machine-generated or synthetic
voices. Despite having access to such technologies, celebrities, politicians, and other well-known individuals are victims.
The main source of concern is a cutting-edge technique known as deepfake, which primarily uses GAN to generate syn-
thetic audio to impersonate actual people. Because of the numerous ways to generate synthetic speech, the challenge of
synthetic speech detection is extremely difficult. Nevertheless, synthetic speech may be created using basic cut-
and-paste methods that accomplish waveform concatenation, which is often accessible as open-source toolkits. It may
also be produced using the voice stream's source or bandpass filter model (K.-D. Xu et al., 2022). Lately, numerous
CNNs-based approaches for generating synthetic audio have been presented. These provide incredibly realistic results
 TABLE 6 Video deepfake techniques' methods, attributes, and features.

Security Simulation Using

Authors Main idea Advantages Research challenges used? Dataset environment TL? Method Usage?
HEIDARI ET AL.

Güera and Delp Presenting a temporal- • High accuracy • High complexity No Deepfake-TIMIT and Python No CNN + Videos
(2018) aware method for • High robustness FaceForensics++ RNN
detecting dataset
automatically.
X. H. Nguyen et al. Using deep 3D CNN • The findings perform • High complexity No FaceForensicsþþ and Python No CNN Videos
(2021) that collects spatio- with binary accuracy • High energy VidTIMIT datasets
temporal rates of 99.4 and 94.5 consumption
characteristics from on the high-quality and
a sequence of brief low-quality data sets,
consecutive frames. respectively
Jung et al. (2020) Providing a method • Accuracy rate of 87.5% • The number of eye No Their deepfake dataset Python No GAN Eye
for detecting blinks was linked blinking
deepfakes created to a mental pattern
by the GANs model. disorder that was
linked to dopamine
activity
Karandikar et al. Using a pre-trained • High accuracy • High complexity No Celeb-DF Dataset Keras for Yes CNN Videos
(2020) CNN to collect face • Low complexity python
characteristics to
extract hidden
features.
Z. Zhao, Zhou et al. Proposing the MFNN • Greater benefit in • High energy No FaceForensics++ TensorFlow + No CNN Videos
(2021) technique for identifying low-quality consumption Keras DL library
capturing artifacts deepfake videos • High complexity
at various levels. • High accuracy
Z. Xu et al. (2021) Providing an SCNN • High accuracy • High complexity No Deepfake-TIMIT Python No CNN Videos
approach for • Low delay • High energy dataset,
detecting face consumption FaceForensics++
modification • Low robustness dataset, and DFDC
techniques Preview dataset
deepfakes.
Kohli and Gupta Developing a • Strong robustness • High energy No FaceForensics++ and MATLAB R2018b No Frequency CNN Videos
(2021) frequency-based consumption Celeb-DF (v2)
CNN method. • High delay dataset
Chen and Tan Proposing feature • Solves the overfitting • High energy No Deepfake-TIMIT + Python Yes CNN + BP-DANN Videos
(2021) transfer, a problem consumption FaceForensics++
technique based on • High complexity
unsupervised
domain adaptation.

(Continues)
21 of 45

19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
 TABLE 6 (Continued)
22 of 45

Security Simulation Using

Authors Main idea Advantages Research challenges used? Dataset environment TL? Method Usage?
A. Yan, Yin-He, Evaluating the frame- • Robust to the • A generic model No Celeb-DF and PyTorch No CNN Videos
et al. (2021) level and compression factor cannot recognize FaceForensics++
temporality-level of several forms of
compressed manipulated facial
deepfake films. videos
Caldelli et al. (2021) Proposing an • Improved robustness in • High energy No FaceForensics++ Python No CNN Videos
approach that a more realistic cross- consumption
makes use of optical forgery operational
flow fields. situation
L. Yan, Yin-He et al. Presenting a • Lower network • Low robustness No FaceForensics, Python No CNN Videos
(2021) lightweight 3D parameters • Low reliability deepfake-TIMIT,
CNN for deepfake • Achieved accuracy DFDC-pre and
detection. with SRM is 98.07 celeb-DF
Mitra et al. (2020) Presenting a CNN- • High accuracy • High complexity No FaceForensics++ Keras with the Yes CNN Videos
based model and a • Low robustness TensorFlow
classifier.
Suratkar et al. Using a CNN • Strong robustness • Security risks were No FaceForensics++, DLIB for python Yes CNN Videos
(2020) architecture and the • Archived accuracy not taken into Google Deep Fake
TL methodology for is 98% account. Dataset, and Deep
exposing such fake Fake Dataset from
videos. Facebook
Bonettini et al. Detecting facial • High accuracy • Moderate delay No FF++ and DFDC Pytorch No CNN Detecting
(2021) alteration in video datasets face alteration
sequences, targeting
both traditional
computer graphics
and DL-produced
fake movies.
Cozzolino et al. Constructing a • High accuracy - No Vision dataset - No CNN Video forensics
(2019) method that • High robustness
improved the
model-related traces
hidden in a video.
HEIDARI ET AL.

19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
HEIDARI ET AL. 23 of 45

that are difficult to distinguish from genuine speech as well as human listeners. The audio anti-spoofing research group
has addressed the more broad topic of synthetic speech synthesis detection. Various methods based on either hand-
crafted or data-driven feature analysis have been presented in this context. Furthermore, with the recent development
of CNN-based algorithms for synthetic audio synthesis, many of the earlier detectors are sure to fail. In this section, we
will look at how to detect fake voices using DL approaches. So, the approach developed by Borrelli et al. (2021) is cen-
tered on short-term and long-term prediction traces. Considering a spoken audio recording, the purpose is to determine
if the speech is synthetic—for example, created by some methods, or genuine, that is, belongs to a real human speaker.
They focused on both closed-set and open-set settings. The suggested approach determines whether the voice is genuine
or synthetic in the closed-set situation. It also recognizes which algorithm was utilized to produce the voice in the case
of synthetic speech. In the open-set scenario, the suggested technique can also detect whether a phony voice was made
using a previously unseen methodology. They integrated a set of characteristics inspired by the speech processing litera-
ture to collect traces from various types of synthetically created voice recordings. They presented, in particular, a set of
characteristics based on the notion of modeling speech as an auto-regressive process. They examined numerous auto-
regressive orders at the same time to construct this feature set. A set of features is derived from the audio under investi-
gation; a supervised classifier is trained to handle the extracted features' classification issue. They offer a set of charac-
teristics based on the notion of modeling speech as an auto-regressive mechanism. This characteristic framework is
defined by taking into account many distinct auto-regressive orders at the same time. Furthermore, they explored the
impact of integrating the suggested characteristics with the bicoherence-based features to see whether they enhance
each other. The suggested features are driven by the widespread use of the source-filter paradigm to analyze and syn-
thesize speech signals. In essence, they suggest extracting various statistics derived by filtering the signal under investi-
gation using short-term and long-term predictors while taking into account different prediction phases. The findings
indicate that their results are more accurate than other solutions. In certain situations, combining all of the qualities is
advantageous.
Also, C.-Z. Yu et al. (2020) proposed a lip-based visual speaker authentication technique to protect against human
imposters as well as deepfake threats. This technique could distinguish deepfake attacks despite previous knowledge of
the video-generation process on the basis that the attackers have insufficient information about the customer and can-
not perfectly recreate the client's talking habit while speaking arbitrary immediate texts. Speaker Authentication
Dynamic Talking Habit Network-based (SA-DTH-Net) is a newer DNN that extracts information about the customer's
specific talking pattern to distinguish the customer's lip image sequence from actual imposters and deepfake forgeries.
By combining all word-level authentication results produced from SA-DTH-Net, the ultimate authentication result for a
speaker reciting a random prompt text may be achieved. Also, the analysis indicates that their strategy can effectively
reject the majority of deepfake attempts created by various manipulation methods. As a result, their technique might be
a viable option for universal deepfake detection in Visual Speaker Authentication (VSA) systems.
In addition, Mittal et al. (2020) demonstrated a learning-based technique for distinguishing between genuine and
phony deepfake multimedia content. They collect and evaluate the similarity between the two auditory and visual
modalities from within the same video to optimize information for training. They also compare and extract effective
clues relating to observed emotion from the two modalities inside a video to determine if it is “genuine” or “fake.” Moti-
vated by the Siamese network architecture and the triplet loss, they suggested a DL network. To validate their model,
they give the Area Under the Curve (AUC) metric on two large-scale deepfake detection datasets, the deepfake-TIMIT
Dataset and DFDC. They evaluate their technique to multiple SOTA deepfake detection approaches, reporting per-
video AUCs of 84.4% on the DFDC dataset and 96.6% on the DF-TIMIT dataset. Their method is the first to use both
audio and video modalities and perceived emotions from two modalities to detect deepfakes.
Besides, R. Wang et al. (2020) introduced DeepSonar, a method that monitors the learned neuron behaviors from a
voice synthesis system to detect AI-generated phony sounds. Their research offers a novel approach to detecting AI-
assisted multimedia forgeries by observing neuron activities to develop a reliable and effective detector. Analyses of the
three datasets show that it is effective and robust and that it has real-world promise in a loud background. They stated
that while creating a detector, robustness should be prioritized because numerous manipulations of voices can be read-
ily disguised as routine activities. In contrast, manipulation of images is restricted and detectable. They use the power
of layer-wise neuron activation patterns to provide a clearer signal to classifiers than raw inputs in the hopes of captur-
ing the tiny variations between real and DL-generated fake voices. Studies on three datasets encompassing both
Chinese and English languages were done to confirm DeepSonar's high detection rates (98.1% average accuracy) and
low false alarm rates (approximately 2% mistake rate) in detecting phony voices. Furthermore, extensive testing results
demonstrate its resistance to manipulation attacks.
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
24 of 45 HEIDARI ET AL.

Finally, Wijethunga et al. (2020) proposed a method for training and validating models to get some of the greatest
accuracy outcomes. Audio signal processing techniques, speaker diarization procedures, and synthetic speech detection
processes are all discussed in their study. Although these signal denoising tasks were completed successfully, the
approaches could be enhanced by using better filtering techniques and enhancing the datasets. The speech-denoising ele-
ment helps clean and preprocesses audio using the Multilayer-Perceptron and CNN architectures, respectively, with 93%
and 94% accuracy. Natural Language Processing (NLP) for text conversion with 93% accuracy and the RNN model for
speaker tagging with 80% accuracy and 0.52 Diarization-Error Rate was used to accomplish speaker diarization. The final
component uses a CNN architecture to discern actual and false sounds accurately. NLP for text conversion with 93% accu-
racy, RNN model for speaker tagging with 80% accuracy, and 0.52 Diarization-Error-Rate was used to accomplish speaker
diarization. The final module takes a CNN architecture to discriminate between real and fraudulent audio with a 94%
accuracy. With these findings, this study will significantly contribute to the field of speech analysis. Also, the speech-
denoising component cleans and preprocesses audio using the Multilayer-Perceptron and CNN architectures, respectively,
with 93% and 94% accuracy. Table 7 lists the deepfake sound applications that use DL techniques and their features.

5.4 | Fake hybrid multimedia detection

Deepfakes (algorithmically modified footage, photos, audio, and videos) could generate huge social upheaval when
combined with the virality of social media. To verify the validity of digital content, anti-disinformation technologies
such as deepfake detection algorithms or immutable metadata are needed. It is critical to have correctly developed tools
to detect, analyze, and defeat deepfake digital content when interacting with multimedia deepfake technology. Multi-
media material could take many forms, including films, artwork, photos, and sound recordings, to mention a few. Even
if there is a trustworthy, secure, and trusted mechanism to trace the history of digital content, achieving this goal can
be difficult. This part will look at how to use DL techniques to recognize hybrid multimedia content. So, Khalil et al.
(2021) presented a deepfake detection approach that combines texture feature extraction with a High-Resolution Net-
work (HRNet)-based strategy for training a capsule network to enhance classification accuracy. The core idea behind
the HRNet would be to prevent the loss of any spatial information, this matched the capsule network's fundamental
idea rather well—except that the original HRNet also had a pooling layer at the very end. The pooling layer from the
HRNet is taken out, and the raw feature vector produced by the HRNet has been used to best suit the two ideas. When
a capsule network (CapsNet) is combined with an improved routing approach (Xiao et al., 2020; Xiao et al., 2021) for
classification, it offers greater feature abstraction and representation capabilities without requiring a lot of data or
parameters. The use of You Only Look Once (YOLO-v3) as a face detector resulted in a significant decrease in false-
positive predictions, eliminating the requirement for additional preprocessing steps. The model was enhanced and
improved as a result of the merging of two distinct in-nature feature extraction techniques. The texture-based method
employing Local Binary Patterns (LBPs) gave an artifacts-specific flavor to their model by counting on the alteration
between the textures of the false face and the surrounding backdrop. CNN-based HRNet supplied informative represen-
tations to their capsule network while maintaining the capsule idea of not sacrificing any spatial information to get the
most out of little details. The HRNet's findings were improved by performing an augmentation–normalization step on
the entered pictures. With its superior feature abstraction and classification performance, CapsNet determined if the
presented picture was real or false.
Also, Chintha et al. (2020) developed the XcepTemporal convolutional RNN system for deepfake detection. They uti-
lized an XceptionNet CNN to identify prominent and efficient face features. This representation is fed through
bidirectional recurrence layers, which help discover temporal discrepancies. They train their model with both standard
cross-entropy and KL divergence loss functions. They offer a companion architecture that obtains audio feature represen-
tations by stacking several convolution modules on the audio side. This audio embedding is also routed through a bidirec-
tional recurrent layer. They show the robustness of their techniques for visual deepfake identification to both compression
and out-of-sample inference on the prominent FaceForensics++ and Celeb-DF datasets, as well as the newly released
deepfake detection challenge. The ASVSpoof 2019 challenge dataset illustrates the resilience of their techniques for audio
spoof detection against compression and out-of-sample inference. Their approaches achieve new benchmark standards for
deepfake visual and spoof audio detection and show that they generalize well to unexpected attacks.
Plus, Kong et al. (2021) suggested a cross-modality approach that uses visual and auditory information to discover
the hidden face behind the deepfake material. Every recovered audio segment is matched with all false faces retrieved
from the related deepfake video, and each fake face is coupled with one matching ground truth face. The pipeline
 TABLE 7 Sound deepfake techniques' methods, properties, and characteristics.
HEIDARI ET AL.

Research Security Simulation Using

Authors Main idea Advantages challenges used? Dataset environment TL? Method Usage?
Borrelli et al. Providing a • High accuracy • High No The ASVSpoof 2019 Scikit-learn— No LSTM AI-synthesized fake voices
(2021) collection of • Low complexity response time logical access audio python library
characteristics • It is difficult to dataset
built on the recognize some
notion of families of
describing the synthetic speech
speech as an recordings in the
auto-regressive open-set
mechanism. situation
C.-Z. Yang Proposing a deep • High accuracy • Not taking into No GRID dataset and Python No Deep CNN Speaker authentication
et al. (2020) CNN-based visual • Low energy account various MOBIO dataset
speaker consumption scenarios
authentication • High delay
method.
Mittal et al. Using the • AUC of 84.4% on • High complexity No TIMIT and DFDC PyAudioAnalysis No RNN + Audio-visual deepfake detection
(2020) similarities the DFDC and • High delay dataset LSTM+
between audio- 96.6% on the DF-
visual modalities TIMIT datasets
and the similarity
between affective
cues.
R. Wang et al. Developing a • Strong robustness • High complexity No FakeOrReal Python No CNN-DNN AI-synthesized fake voices
(2020) method based on • Obtained accuracy • High energy dataset, lab, MC-TTS,
observing neuron is 98.1 consumption and Sprocket-VC
activities.
Wijethunga Suggesting a system • NLP for text • High delay No FakeOrReal TensorFlow Yes CNN + RNN AI-synthesized fake voices
et al. (2020) that can conversion with 93% • High energy dataset + the AMI
differentiate accuracy consumption Corpus dataset
between actual • RNN model for • Low robustness
and synthetic speaker labeling
speech in a group with 80% accuracy
conversation. •
0.52 for
Diarization-
Error-Rate
25 of 45

19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
26 of 45 HEIDARI ET AL.

comprises two parts: (a) an invertible face autoencoder and (b) an audio-visual conditional face reconstruction. They
train a generic face autoencoder to produce representations of ground truth faces with a specific resolution. The
approach is the first of its kind in predicting the genuine face from all available data. The modality transfer is conducted
at the feature level to accurately predict the face appearance, resulting in a promising performance in quantitative and
qualitative assessments. It can have a significant influence on a wide range of applications in multimedia security and
forensics.
Besides, Sun et al. (2020) presented a Landmark Breaker technique to prevent deepfake formation by breaking the
step-facial landmark extraction requirement. Landmark Breaker is considered a new loss function that encourages
the difference between anticipated and original heatmaps and optimizes it with the momentum iterative fast gradient
sign approach. They primarily target CNN-based facial landmark extractors due to their superior performance. To
accomplish it, they introduce adversarial perturbations into the facial landmark extraction process, preventing the input
faces to the deepfake model from being adequately matched. Also, Landmark Breaker has been validated using the
Celeb-DF dataset, demonstrating its efficiency in extracting unpleasant face landmarks. They further look into how
Landmark Breaker performs with different parameter values. Unlike detecting approaches that only operate after a
deepfake has been generated, Landmark Breaker goes one step further and prevents deepfake development.
In addition, Nasar et al. (2020) created a toolkit that uses cutting-edge technology to detect deep fakes in films, pic-
tures, and audio. As a result, their scheme employs a combinational strategy that combines image processing with the
CNN network to identify forgeries in deepfakes. Their system is divided into four components. The first is the data prep-
aration module, which converts the input samples to picture samples. The second module is the data enhancement
module, which removes the image's noise component. Following that is the CNN network, where the model is trained
and tested, and finally, the detection module, which does real-time detection against various media files from multiple
platforms. The model's accuracy was tested against different datasets, including the deepfake-TIMIT dataset, Face
Forensic++, and others, and the accuracy was nearly 0.9 for all three model files.
Chugh et al. (2020) introduced a bimodal deepfake detecting model based on the Modality Dissonance Score (MDS),
it measures the similarity between audio and visual streams for real and fake videos, allowing them to be distinguished.
The MDS is represented using the contrastive loss estimated over segment-level audio-visual data and constraints
authentic audio-visual streams to be closer than their fake counterparts. In addition, the unimodal streams are sub-
jected to cross-entropy loss to ensure that they acquire discriminative features independently. Simulations indicate that
(a) on the DFDC dataset, the MDS-based fake detection framework can attain state-of-the-art performance, and (b) on
top of the contrastive loss, the unimodal cross-entropy losses provide additional benefits to improve fake detection per-
formance. The suggested approach's explainability and interpretability are proven using audio-visual distance distribu-
tions for actual and fraudulent films, Grad-CAM results representing MDS network attention regions, and forgery
localization results.
Finally, Chan et al. (2020) proposed a technique that develops on the experience acquired through the use of various
LSTM-CNN to analyze visual and audio parts of digital media, as well as developing a descriptive caption to go along
with the extracted feature information to form a representative hash that is distinctive to the content. If this material
were not incorporated into a Hyperledger Fabric (HF) framework, deepfakes created from it would lack historical prov-
enance. Permissioned blockchain, particularly HF combined with LSTMs for audio/video/descriptive captions, is a step
in making deepfake media a viable tool. The original artist's verification of untampered data would be required for orig-
inal content. The smart contract integrates several LSTM networks into a method that enables the tracing and tracking
of digital content's historical origin. The outcome is a conceptual model for digital media Proof of Authenticity (PoA)
employing a decentralized blockchain with several LSTMs as deep encoders to create unique discriminative characteris-
tics, that are then compressed and hashed into a transaction. Their research implies that they trust the video at the
point of reception. However, Table 8 lists the deepfake hybrid multimedia applications that use DL techniques and their
features. Also, we will thoroughly examine and analyze the results with all of their features in the next part.

6 | R ESULTS A ND COMPARIS ONS

In the previous section, we looked at four different forms of deepfake detection methods that used different sorts of DL
models. However, public opinion in media information has been eroded by deepfakes, as seeing them no longer equates
to trust in them. They have the power to intensify political tensions, cause public indignation, inspire violence, or even
start a war. They also can boost hate speech and disinformation. This is particularly important today because deepfake
 HEIDARI ET AL.

TABLE 8 Hybrid multimedia deepfake systems' methodologies, traits, and features.

Research Security Simulation Using

Authors Main idea Advantages challenges used? Dataset environment TL? Method Usage?
Kohli and Using LBP to highlight the • Improvement of • High No Deepfake detection Not mentioned No LBP and CNN Image and video
Gupta differences between 20.25% in the area complexity challenge preview detection
(2021) textures of the actual and under the curve • High energy (DFDC-P) dataset
forged parts of the picture. • Very few false consumption
positives
Chintha et al. Presenting a convolutional • Strong robustness • High No FaceForensics++, Keras for Yes CNN Deepfake video
(2020) bidirectional recurrent complexity Celeb-DF, and the python and audio
architecture. • High delay ASVSpoof 2019 detection
logical access audio
datasets
Kong et al. Proposing a method to use • High accuracy • High No VoxCeleb2, CelebDF2, Python Yes Auto-encoders Deepfake video
(2021) learning to produce • Strong robustness complexity and authentic videos and audio
features of the real face. • High delay audios from Youtube detection
Sun et al. Proposing Landmark Breaker • Strong robustness • High energy No Celeb-DF dataset PyTorch No CNN Image and video
(2020) to block the development of • High reliability consumption detection
deepfake by disrupting face
landmark extraction.
Nasar et al. Combining image processing • Obtained accuracy • Low No Deepfake-TIMIT Keras Yes CNN Images, videos,
(2020) with the CNN network to as 0.9 robustness dataset, face and audio
provide a combinational • Low Forensic++
method. reliability
Chugh et al. Using modality dissonance • Increasing the AUC • High delay No DFDC and deepfake- Python No CNN Videos, and audio
(2020) measured across tiny score by up to 7% • Low TIMIT
temporal segments. reliability
Chan et al. Providing a decentralized • High security • High Yes VidTIMIT database Python No LSTM+ Videos, and audio
(2020) blockchain and • Moderate accuracy complexity CNN+
discriminative digital blockchain
media.
27 of 45

19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
28 of 45 HEIDARI ET AL.

techniques are becoming more accessible, and social media platforms can swiftly propagate malicious information.
Also, deepfakes do not always need to be broadcast to a large audience to have a negative impact. People who build
deepfakes for malevolent purposes need to disseminate them to target audiences as part of a sabotaging campaign. They
do not need to use social media to do it. For instance, security agencies could use this strategy to sway influential peo-
ple's judgments, including politicians, posing national and international security risks (Wu et al., 2021; Z. Zhang, Luo,
et al., 2020; Wei Zheng, Xun, Wu, et al., 2021). The scientific community has intended to build deepfake detection algo-
rithms to address the worrying problem of deepfake, and various results have been published. So, this paper gives an
overview of common methodologies as well as an assessment of cutting-edge techniques. A struggle is brewing between
people who utilize powerful ML to build deepfakes and others who attempt to detect them. While the accuracy of
deepfakes improves, so does the effectiveness of detection systems. The idea is that what DL has broken can also be
mended by DL. Detection systems are now in their infancy, and various approaches have indeed been presented and
tested, although on scattered datasets. Creating a constantly updated benchmark dataset of deepfakes to test the ongo-
ing development of detection systems is one way to improve detection method effectiveness. This will make it easier to
train detection algorithms, particularly those based on DL, involving a significant training set. Also, the detection
mistake was seen in the results for all detection systems. As we spoke about in Section 5.4, the research revealed that a
lip-syncing-based method could not identify face-swapping because GANs could indeed generate high-quality facial
expressions that match audio speech, implying that only image-based algorithms could detect deepfake videos with
high accuracy and that potential more modern technologies for face exchanging will be harder to identify. Several of
the prevalent preventative measures indicated across articles included a dependable screening as well as filtering
method between all systems, including such social networks, YouTube, and others, to automate the detection and reme-
diation of fake news, improve legislator and legal requirements, and the use of watermarking tools into devices as then
digital content creates immutable metadata that can be tracked. Although the general subgroup accuracy for detecting
deepfake may vary, it is surprising to see that there is little to no notable difference between deepfake and the actual
image or video.
During the last 2 years, deepfake films have gained widespread notice, representing a great danger to social security.
Towards this purpose, the authors have conducted a large number of investigations and achieved great progress, as we
discussed in Section 5.2. In the previous deepfake datasets, recent detection methods obtain about 100% detection accu-
racy. Nevertheless, with newly constructed datasets, the accuracy of current detection techniques is not perfect. Also,
the average accuracy of detection methods offered in the latest methods competition was just 65%, demonstrating that
existing detection approaches are still far from addressing the needs of practical settings. At the same time, existing
research seeks to identify abstract information using a complicated network structure. Although better detection perfor-
mance is achieved, the improvement in network complexity increases computation costs. The detection impact
vs. network complexity is a tradeoff. A reasonable solution, we suggest, should provide improved precision detection
while reducing network complexity. It is necessary to develop appropriate detection techniques to summarize past algo-
rithms and explore new research approaches in such conditions. Our research study provides a detailed panoramic view
of the domain, which includes up-to-date specifics: types of facial manipulations, public databases for research, facial
manipulation techniques, and benchmarks for detecting each facial manipulation group, which would include crucial
outcomes obtained by much more representative mobsters. In general, most contemporary face manipulations appear
to be straightforward to identify in controlled circumstances, namely when false detectors are tested in the same condi-
tions they were taught. It has been established in the majority of the criteria in this study, with very low mistake rates
in manipulation identification. This situation, nevertheless, would not be very practical because fraudulent pictures
and films are frequently shared on social media, with significant differences in compression level, scaling, noise, and
other factors. Facial manipulation techniques are also getting better all the time. Those certain factors motivate more
studies into the potential of fake detectors to generalize to new situations. Various works have looked at this topic in
the past. A deeper analysis could follow in the footsteps of the most recent articles, as they do not necessitate the use of
false movies for training, allowing for improved generalization of hidden threats.
A solution that aims to make accurate choices on the integrity of videos and photos must have numerous tools to
address most of the relevant operating conditions. In truth, each approach only works when appropriate hypotheses are
met, and when these are not met, the technique becomes entirely useless. In the case of splicing, a tool for copy-move
detection, for example, and will be useless. However, combining various tools is necessary to broaden the range of
detectable forgeries and increase the detection capability for each one. In essence, the traces to be identified are typi-
cally quite weak and could be rapidly camouflaged by deliberate attacks, such as those mentioned in the previous sec-
tion, as well as by ordinary processing processes. As a result, combining numerous techniques designed to identify
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
HEIDARI ET AL. 29 of 45

comparable attacks using various attitudes is likely to increase performance, particularly robustness, in the face of both
benign and malicious disturbances (D. Yang, Zhu, et al., 2022). On the other hand, maximizing the number of clues is a
common investigative technique. A forensic tool's basic methodology is identifying appropriate low-level features from
the original data, analyzing them to get a scalar score or a probability vector, and then evaluating the latter to reach the
final judgment. As indicated in Section 5.4, a combination can occur at any three levels (feature, measurement, or
abstract), each with its own advantages and disadvantages. When many tools are used, dealing at the feature level has
considerable issues, as shown in various studies, especially when there are many features to deal with, and the complex-
ity of establishing representative datasets for all probable important cases. Also, abstract-level integration could have
already destroyed valuable information, diminishing the capacity to exploit cross-tool relationships. Studying at the
level of measurement could be a good middle ground between these two extremes. While integration is unquestionably
a valuable tool for improving performance, it is unclear how to merge disparate data bits effectively. Interestingly, the
fraudulent user went one step further and replicated some content inside the poster, which was identified by a copy-
move-based detector. Nevertheless, several alterations were required to construct a plausible forgery, and so multiple
traces were left. It is not easy to combine these results.
Modern audio generation methods exhibit fingerprint artifacts and repeated inconsistencies across temporal and
spectral domains. As we saw in previous sections, these artifacts could be well captured by frequency domain analysis
over the spectrogram. Mel Frequency Cepstral Coefficients (MFCCs) were used in various speech processing tech-
niques. It has been discovered that extracting features from an audio signal and feeding them into the base model pro-
duces significantly better results than directly feeding the raw audio signal into the model. The result of applying
MFCC is a matrix containing feature vectors extracted from all of the frames. The rows in this output matrix represent
the corresponding frame numbers, and the columns represent the corresponding feature vector coefficients. Finally,
this output matrix is used in the classification process. A spectrogram is a visualization of a signal's frequency spectrum,
where the frequency spectrum of a signal is the frequency range that the signal contains. According to research,
humans do not perceive frequencies on a linear scale, so the Mel scale mimics how the human ear works. A mel-
spectrogram also renders frequencies above a certain threshold logarithmically (the corner frequency). Mel-
spectrograms are generated by applying a Fourier transform to a signal's frequency content and converting it to the
mel-scale, whereas MFCCs are generated by a Discrete Cosine Transform (DCT) into a mel-frequency spectrogram.
The linear audio spectrogram is best suited for applications in which all frequencies are equally important, whereas
mel-spectrograms are better suited for applications in which human hearing perception must be modeled. Data from
mel-spectrograms can also be used in audio classification applications.

6.1 | Analysis of results

However, Section 5 looked into four categories and 32 publications about applying DL approaches in deepfake detection
applications. The most severe flaw in these methods was a lack of comparison across the offered approaches, which is
predictable given the domain's novelty. As a result, identifying flaws in existing approaches might be difficult. Another
point of contention has been the proposed methods' algorithm, which was not explicitly stated; the very least that could
be expected of the proposed methods is a thorough explanation of the algorithms used, which, unfortunately, hardly
any of the plans, algorithms, or pseudo-code did not provide. As a result, we provide a complete overview of the
approaches covered in a variety of domains, including the simulation environment, applications, parameters targeted
by the articles, protection mechanisms, TL methods, datasets, and so on. In addition, Python is the most popular pro-
gramming language for this type of task in the case of a simulation, theoretical, or implementation setting about the
proposed approaches, which is a very enticing component for researchers to employ in future work and has a wide vari-
ety of applications. As shown in Figure 9, 90% of strategies used Python for their simulations, with Keras being the most
popular Python library with 24% usage. MATLAB is a rare environment, accounting for only 3% of all environments.
With 15% usage, Pytorch is in the second position. TensorFlow is in third place with 9%, followed by Scikit-learn, Dlib,
PyAuidoAnalys, and Caffe, each with 3%. Additionally, 30%, with the remainder of the works being done in Python,
but without specifying which library was used. Also, two publications failed to mention their environments. Further-
more, as demonstrated in Figure 10. China (14 articles), the United States (7 articles), India (4 articles), South Korea
(3 articles), and Italy (3 articles) have the most published articles in this area, according to a map of the author's nations
based on their afflation, with a focus on deepfake detection applications. So, the authors from these five countries have
contributed the most to the project.
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
30 of 45 HEIDARI ET AL.

FIGURE 9 The distribution of simulation environments used in the DL-deepfake detection techniques.

FIGURE 10 The most contributed countries by the studied publications are depicted in a geo-chart.

Another investigation of DL applications of deepfake detection in this domain found that 13 articles (40.62%)
employed apps to detect fake videos. Given the global reach and the extent to which these videos could affect peace and
security, one of the advantages of this strategy is that algorithms can find fake videos that humans cannot identify from
authentic ones. The second place jointly and simultaneously goes to image and multimedia fake detection, with seven
articles (21.87%). With five articles (15%), the fake audio attack becomes a big danger to the speaker verification system
and is the third most intriguing issue. This is one of the most important uses of DLs in this sector, although it receives
the least attention. The various uses of DL in the context of deepfake detection are depicted in Figure 11. Section 2 dis-
cusses TL in general. This section explains how TL could aid in mitigating errors in detecting fake content such as
video, images, and sounds. As previously stated, developing enough datasets of fake images or any other deepfakes in a
short period is challenging. As a result, utilizing the benefits of DL to deal with deepfakes is a bit difficult. So, TL could
be useful in this situation. A DL model could be trained using a large-scale benchmark dataset, just like TL, and learned
features can detect deepfakes. Only four studies in our evaluation seek to exploit this TL for diverse purposes in the
realm of deepfake detection. The number of studies is small and the majority of the existing research and experiments
have been utilized to detect fraudulent videos. Many of these activities could be made easier if DL interacts with IoT or
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
HEIDARI ET AL. 31 of 45

FIGURE 11 In the context of deepfake detection, DL application usages.

edge computing (B. Cao, Fan et al., 2021). In some situations, TL could be able to assist. A better system can be given
when the most appropriate algorithm is used for edge computing.
Another significant feature is security, as the information about the person is kept confidential for ethical and legal
reasons (Wei Zheng, Xun, Wu, et al., 2021). As a result, data security must be prioritized while establishing security
structures for these processes. Only one of the procedures analyzed is regarded as a security mechanism for maintaining
security in this area and the confidentiality of deepfake detection mechanisms, which is one of the study's most shock-
ing conclusions. It is the principal disadvantage of the methods under consideration. Ensuring the confidentiality and
security of users' information and processed data could be one of the essential areas for future research. As indicated in
Table 9, the lowest metric evaluated in the articles is security; nevertheless, only one of the offered articles has consid-
ered this crucial element. The accuracy metric is the one that gets the most attention in articles, while the robustness
metric comes in second. The complexity is ranked fourth, and the AUC is placed third. Furthermore, the problem with
articles is that most have one target metric and ignore the others almost entirely. Since deepfake usage is one of the
most dangerous human behaviors ever, with numerous individual and social repercussions, it is desired that additional
research be conducted to address the phenomenon, its ramifications, and complexities, taking into account all poten-
tially dangerous factors' features.
Furthermore, our findings showed that CNN is the most frequently employed method in articles (61%) and is used
in almost every category, particularly in image and video deepfake detection. Furthermore, auto-encoders are the least
used technique in these articles, with a 2.8% usage rate. In addition, Figure 12 depicts the frequency of DL approaches
in deepfake detection applications. As a result, our final analysis focuses on the frequency of datasets utilized in the
publications, which is an important aspect of our assessment which is depicted in Figure 13. Faceforensics is the most
often used dataset (24.4%), followed by CelebA and TIMIT. HFM is ranked last with only one use. In this critical sub-
ject, we hope that our research will aid researchers in creating techniques to avoid and detect deepfake content.

6.2 | Criteria of DL/ML methods

The quality of functions is defined by mathematical metrics that show profitable feedback and analysis of a deepfake
detection method's performance. We have to name a few critical parameters: accuracy, MCC, Confusion Matrix, recall,
precision, and F1 score. As a result, as previously stated, accuracy is the most significant indicator for demonstrating
the fraction of accurately recognized observed to satisfy the predicted observation demand. In the time of combining
total values in a confusion matrix, the True Negative to True Positive rate is exploited. The total quantity of patterns
successfully detected is demonstrated by n, and the entire number of patterns is given by t in this equation.

n
A ¼  100 ð1Þ
t

The given number of exact predictions is indicated by P and the rate of True Positive forecasted compared with the
total positively forecasted. Moreover, STP is the representation of the sum of total true positives, when AFP is the repre-
sentation of total false positives.
 32 of 45

TABLE 9 Considered performance evaluation metrics in the examined articles.

Energy Response
Type Authors Method Security TL consumption time Accuracy Delay Complexity AUC Robustness
Fake image Zhang, Zhao, and Li (2020) CNN • • ✓ • ✓ • • ✓ •
detection Lee et al. (2021) GAN • • • • ✓ • • ✓ •
Guo et al. (2021) CNN • • • ✓ ✓ • • • •
Guarnera et al. (2020) Expectation– • • • • ✓ • • • ✓
Maximization+
CNN
I.-J. Yu et al. (2020) CNN • ✓ • • ✓ • • • ✓
Yang et al. (2021) CNN+ SLIC • • • • ✓ • • • •
Hsu et al. (2020) Pairwise learning • • • • ✓ • • • •
Fake video (Güera & Delp, 2018) CNN + RNN • • • • ✓ • • • ✓
detection Nguyen et al. (2021) CNN • • • • ✓ • • • •
Jung et al. (2020) GAN • • • • ✓ • • • •
Karandikar et al. (2020) CNN • ✓ • • ✓ • • • •
Z. Zhao, Zhou, et al. (2021) CNN • • • • ✓ • • • •
Z. Xu et al. (2021) CNN • • • • ✓ ✓ • • •
Kohli and Gupta (2021) Frequency CNN • • • • • • • • ✓
Chen and Tan (2021) CNN + BP-DNN • ✓ • • • ✓ • • •
A. Yan, Yin-He, et al. (2021) CNN • • • • • • • • ✓
Caldelli et al. (2021) CNN • • • • • • • • ✓
L. Yan, Yin-He, et al. (2021) CNN • • • • • • ✓ • •
Mitra et al. (2020) CNN • ✓ • • ✓ • • • •
Suratkar et al. (2020) CNN • ✓ • • ✓ • • • ✓
Bonettini et al. (2021) CNN • • • • ✓ • • • •
Cozzolino et al. (2019) CNN • • • • ✓ • • • •
Fake sound Borrelli et al. (2021) LSTM • • • • ✓ • ✓ • •
detection Yang et al. (2020) Deep CNN • • ✓ • ✓ • • • •
Mittal et al. (2020) RNN + LSTM • • • • • • • ✓ •
R. Wang et al. (2020) CNN-DNN • • • • ✓ • • • ✓
Wijethunga et al. (2020) CNN + RNN • ✓ • • ✓ • • • •
HEIDARI ET AL.

19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
 TABLE 9 (Continued)

Energy Response
Type Authors Method Security TL consumption time Accuracy Delay Complexity AUC Robustness
HEIDARI ET AL.

Fake Khalil et al. (2021) LBP and CNN • • • • • • • ✓ •

multimedia Chintha et al. (2020) CNN • ✓ • • • • • • ✓
Kong et al. (2021) Auto-encoders • ✓ • • ✓ • • • ✓
Sun et al. (2020) CNN • • • • • • • • ✓
Nasar et al. (2020) CNN • ✓ • • ✓ • • • •
Chugh et al. (2020) CNN • • • • • • • ✓ •
Chan et al. (2020) LSTM+CNN+ Blockchain ✓ • • • ✓ • • • •

Note: The bullets indicate that metrics are not being examined.
33 of 45

19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
34 of 45 HEIDARI ET AL.

FIGURE 12 The utilization of DL approaches and their frequency in selected articles in deepfake detection applications.

FIGURE 13 The frequency of datasets used in the articles.

STP
P¼  100 ð2Þ
STP þ AFP

A criterion of the amount of real positive observations is demonstrated by Re Call named recall which can precisely
forecast. Besides, AFN specifies total false negatives in Equation (3).

STP
Re Call ¼  100 ð3Þ
STP þ AFN

In addition, the F1 score is a total functionality criterion determination of Recall and Precision and representation
of Harmonic achieved by Precision and Recall.

2  Re Call  P
F1score ¼  100 ð4Þ
Re Call þ P

Additionally, functionality matrix measurement which weighs forecasted and real observations is a confusion
matrix that utilizes True Negatives, True positive, False Positives labels, and False Negatives. All true predictions are
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
HEIDARI ET AL. 35 of 45

the total number of Positives and Negatives, so all wrong predictions are the aggregated False Negatives and False
Positives.

STP AFP
ð5Þ
AFN TN

Further, a class that is both true and positive is predicted by True Positives. As well, True Negative are wrong pre-
dictions but are Negative to a clasNegatives are Negative predictions but are incorrect. Also, an individual value func-
tionality demonstrator which encapsulates the total Confusion Matrix is the MCC. It presents a more instructive and
correct result than the F1 score and accuracy in evaluating assortment challenges. If the prediction findings are advan-
tageous in four Confusion Matrix zones, provide a high score.

7 | OPEN ISSUES

In the preceding section, we comprehensively examined the mechanisms under consideration. So, considering the con-
text of modern research on deepfake detection, we have compiled a list of upcoming projects requiring considerable
study addresses.

7.1 | Identity switching

Even though various methods have already been reported in the literature, determining which is the best is undoubt-
edly difficult. This is the result of a variety of circumstances. Initially, general methods are trained for a certain database
and compression level, yielding generally successful performance. Nonetheless, they all demonstrate low generalization
to unobserved circumstances. Furthermore, various measures (e.g., accuracy, AUC, etc.) and empirical validation do
not allow for fair evaluations between studies. All of these factors should be taken into account further to develop in
the area. In addition, we want to emphasize the identification conclusions reached in the most recent deepfake datasets
of the second generation, including such deepfake detection challenges and Celeb-DF. Because when fake detectors
already obtain AUC outcomes near to 100% in datasets of the first generation, such as in the UADFV dataset and
FaceForensics++, they all struggle from such a significant performance degradation on the newer releases, with AUC
outcomes underneath 60% in most case scenarios for the Celeb-DF dataset.

7.2 | Synthetic speech detection

Several circumstances necessitate more research into synthetic speech detection. Because of the wide range of synthetic
speech generation technologies, it is still difficult to recognize some families of synthetic voice tracks in an open-set sit-
uation. Furthermore, the majority of the studies only addressed the logical access synthetic speech recognition problem,
but other articles examined a clean recording of each speech. Future research should investigate what occurs if voice
recordings are damaged by noise, coding, or transmission problems. This situation is especially relevant when authors
consider synthetic voice recordings posted on social media sites or utilized live during phone calls.

7.3 | Interpretability

The interpretability of NN-based methods has become an issue. The NN, being a black box model, cannot offer human-
understandable reasons for its result (Fan et al., 2022; L. Zhou et al., 2023). Nevertheless, the detection method must be
interpretable in real-world forensic settings, or convincing findings will be achieved. Several areas have done signifi-
cantly related studies on interpretability, but research on interpretability in deepfake detecting fields has not developed.
The interpretability of deepfake detection techniques remains a significant issue that must be addressed in the
upcoming.
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
36 of 45 HEIDARI ET AL.

7.4 | Generalization

Generalization is an essential indication for measuring algorithm performance and is frequently used to assess the sys-
tem's performance on unknown datasets. The suggested detection methods primarily focus on supervised learning,
which would be prone to overfitting its datasets. Relevant studies conducted in a few articles have demonstrated that
the generalization performance of the existing detection techniques is still inadequate for cross dataset detection tasks.
In practice, there seem to be significant variations in the choosing of source videos and the post-processing of produced
films, leading to various data sets implying unique distributions. To the authors' knowledge, considerable effort has
been made to improve method generalization. Nevertheless, due to their unique design, these algorithms have intrinsic
faults. Some rely significantly on blending stages, making it impossible to identify artifacts in fully synthetic pictures.
As a result, generality remains a pressing issue that must be addressed.

7.5 | Face synthesis

The majority of contemporary modifications focus on GAN architectures, including Style GAN, which produces
extremely realistic pictures. Nonetheless, most detectors can differentiate between actual and fake pictures with near-
perfect accuracy. It is caused by the fact that fake photos are distinguished using unique GAN fingerprints. But even if
authors could erase the GAN fingerprints or bring multiple noise patterns, still maintain highly realistic synthetic
images? Modern methods have concentrated on this study path, which is a problem even for the finest modification
detection methods.

7.6 | Manipulation of attributes

Because most manipulations are based on GAN structures, the same issue mentioned for face synthesis (GAN finger-
print removal) generally applies here. Furthermore, the paucity of database searches for study and the lack of standard
experimental methods to make meaningful comparisons among experiments is worth noting. A further intriguing
future study subject is the creation of various counter-forensic procedures that we feel have a fundamental right to exist.
Furthermore, technologies to fool forensic detectors introduce a new and fascinating participant into the game, chal-
lenging the detectors of false multimedia material. It is indeed worth noting that specific initializations on the first layer
of a network design have already been utilized in the anti-forensics domain. Consideration of anti-forensic technique
assaults might increase the robustness of forensic detectors. As we observed in previous parts, nearly all existing DL-
based anti-forensic algorithms employ a GAN model; it has been shown to provide good results. Nonetheless, several
techniques for eliminating forensic cues might be investigated, ranging from the creation of appropriate network topol-
ogies to the detailed analysis and elimination of forensic traces using tailored layers and loss functions. Therefore, it is
worth noting that specific initializations on the first layer of a network design have been utilized in the anti-forensics
sector. Consideration of anti-forensic technique assaults might increase the robustness of forensic detectors. We con-
sider that rivalry between forensics and anti-forensics would indeed be good for the progress of both fields and is an
intriguing issue to watch. Also, some methods, such as the fast optimization method (Xi et al., 2021), recursive neural
net (J. Li et al., 2017), and Region-Based Convolutional Neural Network (R-CNN) (Wenfeng Zheng, Yin, Chen,
et al., 2021), can increase the efficiency of the related mechanisms.

7.7 | Changing expressions

To the best of our knowledge, the only public database in expression swap is FaceForensics++, as opposed to identity
swap, which has rapidly expanded with the introduction of better deepfake databases. This dataset is distinguished by
easily detectable visual artifacts, yielding close to 100% AUC resulting in many false detection techniques. We motivate
investigators to create and make public more comprehensive datasets depending on cutting-edge methodologies. Also,
authors can intend to enhance the amount and quality of the HFM dataset in the future by incorporating different face
situations to decrease bias against a particular group and by creating various handcrafted fake facial pictures using
other software programs, including Gimp, Pixlr, and Photoshop. Additional direction for research consideration is by
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
HEIDARI ET AL. 37 of 45

using GANs to supplement manufactured false facial pictures for training or to use TL, a domain adaptation approach,
to increase detection accuracy with a short dataset. Authors could adapt this information to identifying handmade facial
manipulations by saving the knowledge obtained when solving a comparable facial fake detection issue by recognizing
GAN-produced fake faces, deepfakes, and FaceSwap to multiple models. Future research can also provide more effec-
tive 3D CNN structures to better fuse the spatial properties, and a lighter CT module will be developed to extract fea-
tures with fewer parameters.

7.8 | Triplet training

Due to the heterogeneous distribution of datasets, the most difficult challenge for deepfake detection tasks is that gener-
alization performance is insufficient to satisfy the demands of actual scenarios. It is challenging for detection algorithms
to understand the inherent difference between real and fraudulent movies in such conditions. To overcome this issue, a
triplet training approach might be used. In the feature space, triplet training reduces the distance between samples of
the same classification and maximizes the distance between other categories' samples. The triplet training approach, in
particular, guarantees that the distance between samples belonging to distinct groups is greater than the distance
between samples belonging to the same classification. As a result, the optimization objective of triplet training would
be to leverage the inherent difference between actual and false videos, therefore aiding in later classification tasks. In
the domain of face liveness detection, triplet training is being used for domain adaptation tasks, proving the ability of
the triplet training method to discover intrinsic differences between genuine and false films, even though the datasets
have different characteristics.

7.9 | Anti-forensics

Anti-forensic technology has been introduced in response to flaws in present forensic techniques. NNs are frequently
employed in the field of deepfake detection to differentiate counterfeit videos. However, because of intrinsic flaws, NNs
cannot withstand adversarial sample assaults. To that aim, investigators must develop more robust methods to survive
hypothetical laboratory attacks to avoid such attacks in complex real-world settings. The advancement of anti-forensics
technology allows for the prediction of potential assaults and the discovery of flaws in current algorithms, allowing for
the improvement of existing methodologies.

7.10 | Multitask learning

Compared to single-task learning, multitask learning involves doing multiple tasks simultaneously and has been shown
to enhance prediction performance. It has been discovered that doing both forgery location and deepfake recognition
simultaneously time improves accuracy in deepfake identification activities. Multitask learning allows designers to exe-
cute two tasks simultaneously, accounting for losses produced by both tasks and enhancing the model's performance.
Moreover, we demonstrate that forging location is critical in the deepfake detection job. As a result, multitask learning
has a significant potential for further improving deepfake detection.

7.11 | Robustness and resilience

Robustness is a term that is frequently used to assess the effectiveness of detection algorithms when they are subjected
to various degradations. Compressed videos are harder to identify when compared to original videos because they disre-
gard a lot of visual information to achieve a greater compression rate. When confronted with low-quality movies, detec-
tion techniques frequently show a reduction in performance when compared to high-quality videos. Video files may be
subjected to procedures including picture reshaping, rotation, and compression. In those kinds of cases, robustness
would become a crucial feature to consider when building detection systems. Adding a noise layer to the detection net-
work to account for different data degradation conditions is an excellent method to increase resilience. Increasing the
resilience of present detection systems will be critical in the future.
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
38 of 45 HEIDARI ET AL.

7.12 | Response time

Processing time has become a critical factor when used in a real-world scenario. Deepfake detection techniques will also be
tightly integrated on streaming media platforms shortly to mitigate the harmful impact of deepfake videos on social security.
Nevertheless, due to their high time consumption, existing detection algorithms are far from being widely implemented in
actual settings. Given that the movies in practice are far longer than 300 frames, that time consumption falls short of satisfy-
ing the demands of large video detection. In the existing studies on deepfake detection, detection accuracy is considered the
only benchmark, with just a few studies focusing on deepfake detection time consumption. Additional emphasis should be
paid in the future to researching how to build an efficient and high-accuracy detecting technique.

7.13 | Human performance

Even though the dangers of internet deepfake movies are well-known, there seems to be currently no systematic or
quantitative research on the perceptual and psychological components that contribute to their deception. Intriguing
questions remain unanswered; for instance, it is feasible that deepfake films have an uncanny valley; what is the sole
noticeable difference between high-quality deepfake movies and actual videos to human eyes; and what types/aspects
of deepfake videos are considerably more successful in deceiving consumers. It will take proper cooperation between
digital media forensics and perceptual and social psychology academics to answer these problems. Such investigations
are undeniably helpful for both research into detection systems and a greater understanding of the societal impact that
deepfake videos might have.

7.14 | Measures to protect

Nevertheless, considering the speed and accessibility of online media, even with the most advanced forensic tools would
primarily be postmortem in nature, relevant just after DL-generated phony face photographs or videos emerge. In addi-
tion to forensic tools, investigators can develop preventative measures to protect persons from being victims of this kind
of attack. The addition of specially designed patterns known as adversarial perturbations, which are undetectable to
human eyes but can result in detection failures, is one such strategy that academics have lately examined. The reason-
ing behind this is as follows: high-quality DL face synthesis models require a huge number of training face photos, often
in the 100, if not millions, gathered via automatic face identification techniques, referred to as face sets. Adversarial per-
turbations “pollute” a face set, resulting in a small number of actual faces and a large number of non-faces that are of
little or no use as training material for DL face synthesis models. The suggested adversarial perturbation generation
approach can either be implemented as a feature of photo/video sharing platforms before a user uploads their own
photos or movies or as a separate tool that the user can use to edit photos and videos before they are shared online.

7.15 | Blockchain

Identity sovereignty is possible with a permissioned blockchain like HF, which means the original artist retains com-
plete control over their work. To understand the complexities inherent in this proposed system, further in-depth discus-
sion, and implementation outcomes are required. If detection techniques reduce inaccuracy, the system may also attach
detection models as a preprocessing step prior to HF. As DL technology advances, such as the development/
improvement in the standards of DL model creation and compression for archival purposes, we hope to implement
more of the interaction of cutting-edge content-unique hashing methods, integrity methods, security measures, and
widely used blockchains in an effort to ensure that.

7.16 | Deepfakes in other forms

Though face-swapping has been the most well-known deepfake video technique, it is far from the most effective. Face-
swapping deepfake videos, for example, have significant limits when it comes to impersonating someone. According to
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
HEIDARI ET AL. 39 of 45

psychological studies, human face recognition depends heavily on facial shape and hairstyle information. As a result,
the person whose face is to be replaced (the target) must have a comparable facial shape and haircut to the person
whose face is swapped to generate a realistic imitating impression. Furthermore, because the synthetic faces must be
spliced into the original video frame, discrepancies between the generated region and the rest of the original structure
could be significant and challenging to hide. Within those ways, the other two types of deepfake videos, notably head
puppetry and lip-syncing, seem to be more efficient and must therefore be the subject of future deepfake detection
research. During recent decades, techniques for researching complete face synthesis or reenactment have advanced rap-
idly. Even though there have not been as numerous convenient and free public software tools for creating such kinds of
deepfake videos as there are for face-swapping videos, the growing complexity of the generation algorithms could soon
change that. Since this generated region differs from that of face-swapping deepfake films, identification systems that
focus on face-swapping artifacts are unlikely to be useful for such videos. As a result, researchers should work on devel-
oping detecting systems for these types of deepfake movies. Performance analysis, experiments with big datasets, scal-
ability, consistency, and reliability (K. Cao, Wang, et al., 2021) are all planned for the future.

7.17 | GAN-based fake datasets

The GAN approach is the most extensively utilized for image classification and detection. It has recently gained popu-
larity in the medical and healthcare sectors (Lv et al., 2022), and it is one of the most enticing methodologies for investi-
gators. Furthermore, GANs are an intriguing ML approach. Because GANs are generative models, they generate new
data instances comparable to the training data. GANs, for example, may produce images that resemble photos of real
faces, even though the faces do not belong to any actual person. The lack of large datasets and databases of high-quality
pictures for training is one of the most significant obstacles preventing using GAN approaches to generate artificially
produced datasets. According to the paper, GAN techniques can create fake datasets without enormous picture
datasets.

8 | CONCLUSION AND L IMITATIO N

Deepfakes have eroded people's faith in digital content when seeing them no longer equates to believing in them. This
becomes especially important today, even as capabilities for making deepfakes are becoming more accessible, and
online platforms can quickly distribute fake content. People's beliefs and truths can be jeopardized in the absence of
deepfake detection methods. In this regard, such methods can help prevent the spread of fake multimedia content
worldwide while also making it easier for the media to detect them. This work offered a systematic review of the DL
mechanisms for deepfake detection. Before presenting the goal of this research, we addressed the advantages and disad-
vantages of some systematic and peer-reviewed studies about DL-deepfake detection algorithms. Also, the advantages
and disadvantages of each mechanism were explored in four categories based on their applicability. The detecting tools
and platforms for DL-deepfake were also examined. According to articles based on qualitative characteristics, most pub-
lications are assessed based on accuracy, AUC, latency, robustness, and complexity. Meanwhile, some functions, such
as security and delay, go unused. Besides, various programming language libraries are utilized to analyze and imple-
ment the discussed methods, with Keras accounting for 24% of the effort. According to the report, recognizing video
deepfakes is classified by 40% of apps. In the fight against deepfakes, we also hope this research will serve as a valuable
reference for future investigation on DL and deepfake applications. In terms of the prospective findings, applying DL to
process deepfake detection takes a lot of time and effort and tight collaboration between government, business, and aca-
demia. On the other hand, DL has been acclaimed as a wonderful methodology for developing intelligent solutions to
these challenges. The findings of this work could assist in developing deepfake detection-based DL algorithms in real-
world scenarios.
Also, we have encountered several obstacles, including the inaccessibility of non-English articles, which has
prevented us from participating in several development programs. Another limitation of our research is that several
articles we looked at had severe flaws in clear descriptions of the algorithms they utilized. Because this is a brand-new
issue with little maturity in articles and methodologies, constraints such as not comparing the suggested strategy to
other existing methods made it difficult to judge the success of the approaches. Also excluded from the review were
studies that did not directly address deepfake detection. Besides, other articles that were nearly identical to deepfake
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
40 of 45 HEIDARI ET AL.

detection were also removed from consideration. Another issue we ran into was the inaccessibility of numerous articles
published by specific publications.

A U T H O R C ON T R I B U T I O NS
Arash Heidari: Conceptualization (equal); data curation (equal); formal analysis (equal); investigation (equal); meth-
odology (equal); project administration (equal); resources (equal); supervision (equal); validation (equal); visualization
(equal); writing – original draft (equal); writing – review and editing (equal). Nima Jafari Navimipour: Conceptuali-
zation (equal); data curation (equal); formal analysis (equal); investigation (equal); methodology (equal); project admin-
istration (equal); resources (equal); supervision (equal); validation (equal); visualization (equal); writing – original draft
(equal); writing – review and editing (equal). Hasan Dag: Conceptualization (equal); data curation (equal); formal
analysis (equal); investigation (equal); methodology (equal); project administration (equal); resources (equal); supervi-
sion (equal); validation (equal); visualization (equal); writing – original draft (equal); writing – review and editing
(equal). Mehmet Unal: Investigation (equal); methodology (equal); resources (equal); supervision (equal); validation
(equal); writing – review and editing (equal).

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.

DATA AVAILABILITY STATEMENT

The article contains all of the data.

ORCID
Arash Heidari https://orcid.org/0000-0003-4279-8551

R EL ATE D WIR Es AR TI CL ES
Deepfake attribution: On the source identification of artificially generated images

R EF E RE N C E S
Afchar, D., Nozick, V., Yamagishi, J., & Echizen, I. (2018). Mesonet: A compact facial video forgery detection network. Paper presented at the
2018 IEEE International Workshop on Information Forensics and Security (WIFS).
Ahmed, I., Ahmad, M., Rodrigues, J. J., & Jeon, G. (2021). Edge computing-based person detection system for top view surveillance: Using
CenterNet with transfer learning. Applied Soft Computing, 107, 107489.
Ahmed, S. R. A., & Sonuç, E. (2021). Deepfake detection using rationale-augmented convolutional neural network. Applied Nanoscience, 13,
1485–1493.
Akhtar, Z., Mouree, M. R., & Dasgupta, D. (2020). Utility of deep learning features for facial attributes manipulation detection. Paper pres-
ented at the 2020 IEEE International Conference on Humanized Computing and Communication with Artificial Intelligence (HCCAI).
Albahar, M., & Almalki, J. (2019). Deepfakes: Threats and countermeasures systematic review. Journal of Theoretical and Applied Informa-
tion Technology, 97(22), 3242–3250.
Aversano, L., Bernardi, M. L., Cimitile, M., & Pecori, R. (2021). A systematic review on deep learning approaches for IoT security. Computer
Science Review, 40, 100389.
Balaji, T., Annavarapu, C. S. R., & Bablani, A. (2021). Machine learning algorithms for social media analysis: A survey. Computer Science
Review, 40, 100395.
Baygin, M., Yaman, O., Baygin, N., & Karakose, M. (2022). A blockchain-based approach to smart cargo transportation using UHF RFID.
Expert Systems with Applications, 188, 116030.
Bekci, B., Akhtar, Z., & Ekenel, H. K. (2020). Cross-dataset face manipulation detection. Paper presented at the 2020 28th Signal Processing
and Communications Applications Conference (SIU).
Biswas, A., Bhattacharya, D., & Kakelli, A. K. (2021). DeepFake detection using 3D-Xception net with discrete Fourier transformation.
Journal of Information Systems and Telecommunication, 3(35), 161–168.
Bonettini, N., Cannas, E. D., Mandelli, S., Bondi, L., Bestagini, P., & Tubaro, S. (2021). Video face manipulation detection through ensemble
of CNNS. Paper presented at the 2020 25th International Conference on Pattern Recognition (ICPR).
Borrelli, C., Bestagini, P., Antonacci, F., Sarti, A., & Tubaro, S. (2021). Synthetic speech detection through short-term and long-term predic-
tion traces. EURASIP Journal on Information Security, 2021(1), 1–14.
Burroughs, S. J., Gokaraju, B., Roy, K., & Khoa, L. (2020). DeepFakes detection in videos using feature engineering techniques in deep learn-
ing convolution neural network frameworks. Paper presented at the 2020 IEEE Applied Imagery Pattern Recognition Workshop (AIPR).
Caldelli, R., Galteri, L., Amerini, I., & Del Bimbo, A. (2021). Optical flow based CNN for detection of unlearnt deepfake manipulations. Pat-
tern Recognition Letters, 146, 31–37.
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
HEIDARI ET AL. 41 of 45

Cao, B., Fan, S., Zhao, J., Tian, S., Zheng, Z., Yan, Y., & Yang, P. (2021). Large-scale many-objective deployment optimization of edge servers.
IEEE Transactions on Intelligent Transportation Systems, 22(6), 3841–3849.
Cao, K., Wang, B., Ding, H., Lv, L., Tian, J., Hu, H., & Gong, F. (2021). Achieving reliable and secure communications in wireless-powered
NOMA systems. IEEE Transactions on Vehicular Technology, 70(2), 1978–1983.
Castillo Camacho, I., & Wang, K. (2021). A comprehensive review of deep-learning-based methods for image forensics. Journal of Imaging,
7(4), 69.
Chan, C. C. K., Kumar, V., Delaney, S., & Gochoo, M. (2020). Combating deepfakes: Multi-LSTM and blockchain as proof of authenticity for
digital media. Paper presented at the 2020 IEEE/ITU International Conference on Artificial Intelligence for Good (AI4G).
Chen, B., & Tan, S. (2021). FeatureTransfer: Unsupervised domain adaptation for cross-domain deepfake detection. Security and Communi-
cation Networks, 2021, 1–8.
Chi, H., Maduakor, U., Alo, R., & Williams, E. (2020). Integrating deepfake detection into cybersecurity curriculum. Paper presented at the
Proceedings of the Future Technologies Conference.
Chintha, A., Thai, B., Sohrawardi, S. J., Bhatt, K., Hickerson, A., Wright, M., & Ptucha, R. (2020). Recurrent convolutional structures for
audio spoof and video deepfake detection. IEEE Journal of Selected Topics in Signal Processing, 14(5), 1024–1037.
Chugh, K., Gupta, P., Dhall, A., & Subramanian, R. (2020). Not made for each other-audio-visual dissonance-based deepfake detection and
localization. Paper presented at the Proceedings of the 28th ACM International Conference on Multimedia.
Chung, J., Gulcehre, C., Cho, K., & Bengio, Y. (2015). Gated feedback recurrent neural networks. Paper presented at the International confer-
ence on machine learning.
Cozzolino, D., Poggi, G., & Verdoliva, L. (2019). Extracting camera-based fingerprints for video forensics. Paper Presented at the Proceedings
of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops.
Deshmukh, A., & Wankhade, S. B. (2021). Deepfake detection approaches using deep learning: A systematic review. In V. E. Balas, V. B.
Semwal, A. Khandare, & Patil, M. (Eds.), Intelligent Computing and Networking. Lecture Notes in Networks and Systems, Vol 146.
Springer. https://doi.org/10.1007/978-981-15-7421-4_27
Dixit, P., & Silakari, S. (2021). Deep learning algorithms for cybersecurity applications: A technological and status review. Computer Science
Review, 39, 100317.
Doewes, R. I., Gharibian, G., Zadeh, F. A., Zaman, B. A., Vahdat, S., & Akhavan-Sigari, R. (2023). An updated systematic review on the
effects of aerobic exercise on human blood lipid profile. Current Problems in Cardiology, 48(5), 101108. https://doi.org/10.1016/j.
cpcardiol.2022.101108
Du, M., Pentyala, S., Li, Y., & Hu, X. (2020). Towards generalizable deepfake detection with locality-aware autoencoder. Paper presented at
the Proceedings of the 29th ACM International Conference on Information & Knowledge Management.
Dwivedi, Y. K., Hughes, L., Ismagilova, E., Aarts, G., Coombs, C., Crick, T., & Eirug, A. (2021). Artificial Intelligence (AI): Multidisciplinary
perspectives on emerging challenges, opportunities, and agenda for research, practice and policy. International Journal of Information
Management, 57, 101994.
Esmailiyan, M., Amerizadeh, A., Vahdat, S., Ghodsi, M., Doewes, R. I., & Sundram, Y. (2021). Effect of different types of aerobic exercise on
individuals with and without hypertension: An updated systematic review. Current Problems in Cardiology, 101034. https://doi.org/10.
1016/j.cpcardiol.2021.101034
Fan, Q., Zhang, Z., & Huang, X. (2022). Parameter conjugate gradient with secant equation based Elman neural network and its convergence
analysis. Advanced Theory and Simulations, 5(9), 2200047.
Feng, Y., Zhang, B., Liu, Y., Niu, Z., Dai, B., Fan, Y., & Chen, X. (2021). A 200–225-GHz manifold-coupled multiplexer utilizing metal wave-
guides. IEEE Transactions on Microwave Theory and Techniques, 69(12), 5327–5333.
Fernandes, S. L., & Jha, S. K. (2020). Adversarial attack on deepfake detection using RL based texture patches. Paper presented at the
European Conference on Computer Vision.
Fink, O., Wang, Q., Svensen, M., Dersin, P., Lee, W.-J., & Ducoffe, M. (2020). Potential, challenges and future directions for deep learning in
prognostics and health management applications. Engineering Applications of Artificial Intelligence, 92, 103678.
Garg, A., & Mago, V. (2021). Role of machine learning in medical research: A survey. Computer Science Review, 40, 100370.
Gosse, C., & Burkell, J. (2020). Politics and porn: How news media characterizes problems presented by deepfakes. Critical Studies in Media
Communication, 37(5), 497–511.
Guarnera, L., Giudice, O., & Battiato, S. (2020). Fighting deepfake by exposing the convolutional traces on images. IEEE Access, 8, 165085–
165098.
Güera, D., & Delp, E. J. (2018). Deepfake video detection using recurrent neural networks. Paper presented at the 2018 15th IEEE Interna-
tional Conference on Advanced Video and Signal Based Surveillance (AVSS).
Guo, Z., Yang, G., Chen, J., & Sun, X. (2021). Fake face detection via adaptive manipulation traces extraction network. Computer Vision and
Image Understanding, 204, 103170.
Habeeba, M. S., Lijiya, A., & Chacko, A. M. (2021). Detection of Deepfakes using visual artifacts and neural network classifier. In Innovations
in electrical and electronic engineering (pp. 411–422). Springer.
Heidari, A., Jafari Navimipour, N., Unal, M., & Toumaj, S. (2022). Machine learning applications for COVID-19 outbreak management.
Neural Computing and Applications, 34, 15313–15348. https://doi.org/10.1007/s00521-022-07424-w
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
42 of 45 HEIDARI ET AL.

Heidari, A., Navimipour, N. J., & Unal, M. (2022). Applications of ML/DL in the management of smart cities and societies based on new
trends in information technologies: A systematic literature review. Sustainable Cities and Society, 104089. https://doi.org/10.1016/j.scs.
2022.104089
Heidari, A., Navimipour, N. J., Unal, M., & Toumaj, S. (2022). The COVID-19 epidemic analysis and diagnosis using deep learning: A system-
atic literature review and future directions. Computers in Biology and Medicine, 141, 105141.
Heidari, A., Toumaj, S., Navimipour, N. J., & Unal, M. (2022). A privacy-aware method for COVID-19 detection in chest CT images using
lightweight deep conventional neural network and blockchain. Computers in Biology and Medicine, 145, 105461.
Hinton, G., Vinyals, O., & Dean, J. (2015). Distilling the knowledge in a neural network. arXiv Preprint arXiv:1503.02531, 2(7).
Hong, T., Guo, S., Jiang, W., & Gong, S. (2021). Highly selective frequency selective surface with ultrawideband rejection. IEEE Transactions
on Antennas and Propagation, 70(5), 3459–3468.
Hongmeng, Z., Zhiqiang, Z., Lei, S., Xiuqing, M., & Yuehan, W. (2020). A detection method for DeepFake hard compressed videos based on
super-resolution reconstruction using CNN. Paper presented at the Proceedings of the 2020 4th High Performance Computing and Cluster
Technologies Conference & 2020 3rd International Conference on Big Data and Artificial Intelligence.
Hossain, M. S., Cucchiara, R., Muhammad, G., Tob on, D. P., & Saddik, A. E. (2022). Special section on AI-empowered multimedia data ana-
lytics for smart healthcare. ACM Transactions on Multimedia Computing, Communications and Applications, 18, 1–2.
Hsu, C.-C., Zhuang, Y.-X., & Lee, C.-Y. (2020). Deep fake image detection based on pairwise learning. Applied Sciences, 10(1), 370.
Hung, J. C., & Chang, J.-W. (2021). Multi-level transfer learning for improving the performance of deep neural networks: Theory and prac-
tice from the tasks of facial emotion recognition and named entity recognition. Applied Soft Computing, 109, 107491.
Iqbal, T., & Qureshi, S. (2022). The survey: Text generation models in deep learning. Journal of King Saud University-Computer and Informa-
tion Sciences, 34(6), 2515–2528.
Jafar, M. T., Ababneh, M., Al-Zoube, M., & Elhassan, A. (2020). Forensics and analysis of deepfake videos. Paper presented at the 2020 11th
International Conference on Information and Communication Systems (ICICS).
Jeyaraj, A., & Dwivedi, Y. K. (2020). Meta-analysis in information systems research: Review and recommendations. International Journal of
Information Management, 55, 102226.
Jia, T., Cai, C., Li, X., Luo, X., Zhang, Y., & Yu, X. (2022). Dynamical community detection and spatiotemporal analysis in multilayer spatial
interaction networks using trajectory data. International Journal of Geographical Information Science, 36, 1–22.
Jiang, J., Li, B., Wei, B., Li, G., Liu, C., Huang, W., & Yu, M. (2021). FakeFilter: A cross-distribution Deepfake detection system with domain
adaptation. Journal of Computer Security, 29(4), 1–19.
Jung, T., Kim, S., & Kim, K. (2020). DeepVision: Deepfakes detection using human eye blinking pattern. IEEE Access, 8, 83144–83154.
Karandikar, A., Deshpande, V., Singh, S., Nagbhidkar, S., & Agrawal, S. (2020). Deepfake video detection using convolutional neural net-
work. International Journal of Advanced Trends in Computer Science and Engineering, 9(2), 1311–1315.
Katarya, R., & Lal, A. (2020). A study on combating emerging threat of deepfake weaponization. Paper presented at the 2020 Fourth Interna-
tional Conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud) (I-SMAC).
Khalil, S. S., Youssef, S. M., & Saleh, S. N. (2021). iCaps-Dfake: An integrated capsule-based model for Deepfake image and video detection.
Future Internet, 13(4), 93.
Kietzmann, J., Lee, L. W., McCarthy, I. P., & Kietzmann, T. C. (2020). Deepfakes: Trick or treat? Business Horizons, 63(2), 135–146.
Kim, E., & Cho, S. (2021). Exposing fake faces through deep neural networks combining content and trace feature extractors. IEEE Access, 9,
123493–123503.
Kohli, A., & Gupta, A. (2021). Detecting DeepFake, FaceSwap and Face2Face facial forgeries using frequency CNN. Multimedia Tools and
Applications, 80(12), 18461–18478.
Kong, C., Chen, B., Yang, W., Li, H., Chen, P., & Wang, S. (2021). Appearance matters, so does audio: Revealing the hidden face via cross-
modality transfer. IEEE Transactions on Circuits and Systems for Video Technology, 32(1), 423–436.
Lee, S., Tariq, S., Shin, Y., & Woo, S. S. (2021). Detecting handcrafted facial image manipulations and GAN-generated facial images using
shallow-FakeFaceNet. Applied Soft Computing, 105, 107256.
Lewis, J. K., Toubal, I. E., Chen, H., Sandesera, V., Lomnitz, M., Hampel-Arias, Z., & Palaniappan, K. (2020). Deepfake video detection based
on spatial, spectral, and temporal inconsistencies using multimodal deep learning. Paper presented at the 2020 IEEE Applied Imagery Pat-
tern Recognition Workshop (AIPR).
Li, A., Masouros, C., Swindlehurst, A. L., & Yu, W. (2021). 1-bit massive MIMO transmission: Embracing interference with symbol-level
precoding. IEEE Communications Magazine, 59(5), 121–127.
Li, A., Spano, D., Krivochiza, J., Domouchtsidis, S., Tsinos, C. G., Masouros, C., & Ottersten, B. (2020). A tutorial on interference exploitation
via symbol-level precoding: Overview, state-of-the-art and future directions. IEEE Communications Surveys & Tutorials, 22(2), 796–839.
Li, D., Ge, S. S., & Lee, T. H. (2020). Fixed-time-synchronized consensus control of multiagent systems. IEEE Transactions on Control of
Network Systems, 8(1), 89–98.
Li, J., Xu, K., Chaudhuri, S., Yumer, E., Zhang, H., & Guibas, L. (2017). Grass: Generative recursive autoencoders for shape structures. ACM
Transactions on Graphics, 36(4), 1–14.
Li, S., Liu, C. H., Lin, Q., Wen, Q., Su, L., Huang, G., & Ding, Z. (2020). Deep residual correction network for partial domain adaptation.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7), 2329–2344.
Li, Y., Zhang, C., Sun, P., Ke, L., Ju, Y., Qi, H., & Lyu, S. (2021). DeepFake-o-meter: An open platform for DeepFake detection. Paper pres-
ented at the 2021 IEEE Security and Privacy Workshops (SPW).
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
HEIDARI ET AL. 43 of 45

Liu, Q., & Celebi, N. (2021). Large feature mining and deep learning in multimedia forensics. Paper presented at the Proceedings of the 2021
ACM Workshop on Security and Privacy Analytics.
Liu, R., Wang, X., Lu, H., Wu, Z., Fan, Q., Li, S., & Jin, X. (2021). SCCGAN: Style and characters inpainting based on CGAN. Mobile Networks
and Applications, 26(1), 3–12.
Luo, G., Yuan, Q., Li, J., Wang, S., & Yang, F. (2022). Artificial intelligence powered mobile networks: From cognition to decision. IEEE
Network, 36(3), 136–144.
Luo, G., Zhang, H., Yuan, Q., Li, J., & Wang, F.-Y. (2022). ESTNet: Embedded spatial-temporal network for modeling traffic flow dynamics.
IEEE Transactions on Intelligent Transportation Systems, 23, 19201–19212.
Lv, Z., Chen, D., Feng, H., Zhu, H., & Lv, H. (2021). Digital twins in unmanned aerial vehicles for rapid medical resource delivery in
epidemics. IEEE Transactions on Intelligent Transportation Systems, 23(12), 25106–25114.
Lv, Z., Li, Y., Feng, H., & Lv, H. (2021). Deep learning for security in digital twins of cooperative intelligent transportation systems. IEEE
Transactions on Intelligent Transportation Systems, 22, 4281–4290.
Lv, Z., Qiao, L., & You, I. (2020). 6G-enabled network in box for internet of connected vehicles. IEEE Transactions on Intelligent Transporta-
tion Systems, 22(8), 5275–5282.
Lv, Z., Yu, Z., Xie, S., & Alamri, A. (2022). Deep learning-based smart predictive evaluation for interactive multimedia-enabled smart
healthcare. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 18(1s), 1–20.
Marei, M., El Zaatari, S., & Li, W. (2021). Transfer learning enabled convolutional neural networks for estimating health state of cutting
tools. Robotics and Computer-Integrated Manufacturing, 71, 102145.
Masi, I., Killekar, A., Mascarenhas, R. M., Gurudatt, S. P., & AbdAlmageed, W. (2020). Two-branch recurrent network for isolating deepfakes
in videos. Paper presented at the European conference on Computer Vision.
Matern, F., Riess, C., & Stamminger, M. (2019). Exploiting visual artifacts to expose deepfakes and face manipulations. Paper presented at the
2019 IEEE Winter Applications of Computer Vision Workshops (WACVW).
Mehta, V., Gupta, P., Subramanian, R., & Dhall, A. (2021). FakeBuster: A DeepFakes detection tool for video conferencing scenarios. Paper
presented at the 26th International Conference on Intelligent User Interfaces.
Meskys, E., Liaudanskas, A., Kalpokiene, J., & Jurcys, P. (2020). Regulating deep fakes: Legal and ethical considerations. Journal of Intellec-
tual Property Law & Practice, 15(1), 24–31.
Mirsky, Y., & Lee, W. (2021). The creation and detection of deepfakes: A survey. ACM Computing Surveys (CSUR), 54(1), 1–41.
Mitra, A., Mohanty, S. P., Corcoran, P., & Kougianos, E. (2020). A novel machine learning based method for deepfake video detec-
tion in social media. Paper presented at the 2020 IEEE International Symposium on Smart Electronic Systems (iSES) (Formerly
iNiS).
Mittal, T., Bhattacharya, U., Chandra, R., Bera, A., & Manocha, D. (2020). Emotions don't lie: An audio-visual deepfake detection method
using affective cues. Paper presented at the Proceedings of the 28th ACM International Conference on Multimedia.
Muhammad, G., & Hossain, M. S. (2022). Light deep models for cognitive computing in intelligent transportation systems. IEEE Transactions
on Intelligent Transportation Systems, 24(1), 1144–1152.
Nasar, B. F., Sajini, T., & Lason, E. R. (2020). Deepfake detection in media files-audios, images and videos. Paper presented at the 2020 IEEE
Recent Advances in Intelligent Computational Systems (RAICS).
Nguyen, H. H., Yamagishi, J., & Echizen, I. (2019). Capsule-forensics: Using capsule networks to detect forged images and videos. Paper
presented at the ICASSP 2019–2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP).
Nguyen, T. T., Nguyen, C. M., Nguyen, D. T., Nguyen, D. T., & Nahavandi, S. (2019). Deep learning for deepfakes creation and detection: A
survey. arXiv Preprint arXiv:1909.11573.
Nguyen, X. H., Tran, T. S., Nguyen, K. D., & Truong, D.-T. (2021). Learning spatio-temporal features to detect manipulated facial videos cre-
ated by the Deepfake techniques. Forensic Science International: Digital Investigation, 36, 301108.
Niknejad, N., Ismail, W. B., Mardani, A., Liao, H., & Ghani, I. (2020). A comprehensive overview of smart wearables: The state of the art lit-
erature, recent advances, and future challenges. Engineering Applications of Artificial Intelligence, 90, 103529.
Nirkin, Y., Wolf, L., Keller, Y., & Hassner, T. (2021). DeepFake detection based on discrepancies between faces and their context. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 44(10), 6111–6121.
Niu, Z., Zhang, B., Dai, B., Zhang, J., Shen, F., Hu, Y., & Zhang, Y. (2022). 220 GHz multi circuit integrated front end based on solid-state cir-
cuits for high speed communication system. Chinese Journal of Electronics, 31(3), 569–580.
Pashine, S., Mandiya, S., Gupta, P., & Sheikh, R. (2021). Deep fake detection: Survey of facial manipulation detection solutions. arXiv
Preprint arXiv:2106.12605.
Pokroy, A. A., & Egorov, A. D. (2021). EfficientNets for DeepFake detection: Comparison of pretrained models. Paper presented at the 2021
IEEE Conference of Russian Young Researchers in Electrical and Electronic Engineering (ElConRus).
Rana, M. S., Nobi, M. N., Murali, B., & Sung, A. H. (2022). Deepfake detection: A systematic literature review. IEEE Access, 10, 25494–25513.
Rana, M. S., & Sung, A. H. (2020). Deepfakestack: A deep ensemble-based learning technique for deepfake detection. Paper presented at the
2020 7th IEEE International Conference on Cyber Security and Cloud Computing (CSCloud)/2020 6th IEEE International Conference on
Edge Computing and Scalable Cloud (EdgeCom).
Saif, S., & Tehseen, S. (2022). Deepfake videos: Synthesis and detection techniques–A survey. Journal of Intelligent & Fuzzy Systems, 42(4),
2989–3009.
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
44 of 45 HEIDARI ET AL.

Sanghvi, B., Shelar, H., Pandey, M., & Sisodia, J. (2021). Detection of machine generated multimedia elements using deep learning. Paper
presented at the 2021 5th International Conference on Computing Methodologies and Communication (ICCMC).
Schlett, T., Rathgeb, C., & Busch, C. (2021). Deep learning-based single image face depth data enhancement. Computer Vision and Image
Understanding, 210, 103247.
Shelke, N. A., & Kasana, S. S. (2021). A comprehensive survey on passive techniques for digital video forgery detection. Multimedia Tools
and Applications, 80, 6247–6310.
Siegel, D., Kraetzer, C., Seidlitz, S., & Dittmann, J. (2021). Media forensics considerations on DeepFake detection with hand-crafted features.
Journal of Imaging, 7(7), 108.
Singh, P., Masud, M., Hossain, M. S., Kaur, A., Muhammad, G., & Ghoneim, A. (2021). Privacy-preserving serverless computing using feder-
ated learning for smart grids. IEEE Transactions on Industrial Informatics, 18(11), 7843–7852.
Soares, M. A. C., & Parreiras, F. S. (2020). A literature review on question answering techniques, paradigms and systems. Journal of King
Saud University-Computer and Information Sciences, 32(6), 635–646.
Sun, P., Li, Y., Qi, H., & Lyu, S. (2020). Landmark breaker: Obstructing DeepFake by disturbing landmark extraction. Paper presented at the
2020 IEEE International Workshop on Information Forensics and Security (WIFS).
Suratkar, S., Johnson, E., Variyambat, K., Panchal, M., & Kazi, F. (2020). Employing transfer-learning based CNN architectures to enhance
the generalizability of deepfake detection. Paper presented at the 2020 11th International Conference on Computing, Communication and
Networking Technologies (ICCCNT).
Swathi, P., & Sk, S. (2021). DeepFake creation and detection: A survey. Paper presented at the 2021 Third International Conference on Inven-
tive Research in Computing Applications (ICIRCA).
Tariq, S., Lee, S., & Woo, S. (2021). One detector to rule them all: Towards a general deepfake attack detection framework. Paper presented at
the Proceedings of the Web Conference 2021.
Tjon, E., Moh, M., & Moh, T.-S. (2021). Eff-YNet: A Dual Task Network for DeepFake Detection and Segmentation. Paper presented at the
2021 15th International Conference on Ubiquitous Information Management and Communication (IMCOM).
Tolosana, R., Vera-Rodriguez, R., Fierrez, J., Morales, A., & Ortega-Garcia, J. (2020). Deepfakes and beyond: A survey of face manipulation
and fake detection. Information Fusion, 64, 131–148.
Tu, Y., Liu, Y., & Li, X. (2021). Deepfake video detection by using convolutional gated recurrent unit. Paper presented at the 2021 13th Inter-
national Conference on Machine Learning and Computing.
Vahdat, S., & Shahidi, S. (2020). D-dimer levels in chronic kidney illness: a comprehensive and systematic literature review. Proceedings of
the National Academy of Sciences, India Section B: Biological Sciences, 1-18.
Verdoliva, L. (2020). Media forensics and deepfakes: An overview. IEEE Journal of Selected Topics in Signal Processing, 14(5), 910–932.
Wang, R., Juefei-Xu, F., Huang, Y., Guo, Q., Xie, X., Ma, L., & Liu, Y. (2020). Deepsonar: Towards effective and robust detection of AI-
synthesized fake voices. Paper presented at the Proceedings of the 28th ACM International Conference on Multimedia.
Wang, Y., Wang, H., Zhou, B., & Fu, H. (2021). Multi-dimensional prediction method based on Bi-LSTMC for ship roll. Ocean Engineering, 242, 110106.
Wang, Z., Ramamoorthy, R., Xi, X., & Namazi, H. (2022). Synchronization of the neurons coupled with sequential developing electrical and
chemical synapses. Mathematical Biosciences and Engineering, 19(2), 1877–1890.
Wang, Z., Ramamoorthy, R., Xi, X., Rajagopal, K., Zhang, P., & Jafari, S. (2022). The effects of extreme multistability on the collective dynam-
ics of coupled memristive neurons. The European Physical Journal Special Topics, 231, 1–8.
Weerawardana, M., & Fernando, T. (2021). Deepfakes detection methods: A literature survey. Paper presented at the 2021 10th International
Conference on Information and Automation for Sustainability (ICIAfS).
Wijethunga, R., Matheesha, D., Al Noman, A., De Silva, K., Tissera, M., & Rupasinghe, L. (2020). Deepfake audio detection: A deep learning
based solution for group conversations. Paper presented at the 2020 2nd International Conference on Advancements in Computing (ICAC).
Wu, X., Zheng, W., Xia, X., & Lo, D. (2021). Data quality matters: A case study on data label correctness for security bug report prediction.
IEEE Transactions on Software Engineering, 48(7), 2541–2556.
Xi, Y., Jiang, W., Wei, K., Hong, T., Cheng, T., & Gong, S. (2021). Wideband RCS reduction of microstrip antenna array using coding meta-
surface with low Q resonators and fast optimization method. IEEE Antennas and Wireless Propagation Letters, 21(4), 656–660.
Xiao, Y., Zhang, Y., Kaku, I., Kang, R., & Pan, X. (2021). Electric vehicle routing problem: A systematic review and a new comprehensive
model with nonlinear energy recharging and consumption. Renewable and Sustainable Energy Reviews, 151, 111567.
Xiao, Y., Zuo, X., Huang, J., Konak, A., & Xu, Y. (2020). The continuous pollution routing problem. Applied Mathematics and Computation,
387, 125072.
Xie, B., Li, S., Lv, F., Liu, C. H., Wang, G., & Wu, D. (2023). A collaborative alignment framework of transferable knowledge extraction for
unsupervised domain adaptation. IEEE Transactions on Knowledge and Data Engineering, 35(7), 6518–6533.
Xie, D., Chatterjee, P., Liu, Z., Roy, K., & Kossi, E. (2020). DeepFake detection on publicly available datasets using modified AlexNet. Paper
presented at the 2020 IEEE Symposium Series on Computational Intelligence (SSCI).
Xu, K.-D., Weng, X., Li, J., Guo, Y.-J., Wu, R., Cui, J., & Chen, Q. (2022). 60-GHz third-order on-chip bandpass filter using GaAs pHEMT
technology. Semiconductor Science and Technology, 37(5), 055004.
Xu, Z., Liu, J., Lu, W., Xu, B., Zhao, X., Li, B., & Huang, J. (2021). Detecting facial manipulated videos based on set convolutional neural net-
works. Journal of Visual Communication and Image Representation, 77, 103119.
Yan, A., Fan, Z., Ding, L., Cui, J., Huang, Z., Wang, Q., & Wen, X. (2021). Cost-effective and highly reliable circuit-components design for
safety-critical applications. IEEE Transactions on Aerospace and Electronic Systems, 58(1), 517–529.
 19424795, 2024, 2, Downloaded from https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1520, Wiley Online Library on [24/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
HEIDARI ET AL. 45 of 45

Yan, J., Jiao, H., Pu, W., Shi, C., Dai, J., & Liu, H. (2022). Radar sensor network resource allocation for fused target tracking: A brief review.
Information Fusion, 86, 104–115.
Yan, L., Yin-He, S., Qian, Y., Zhi-Yu, S., Chun-Zi, W., & Zi-Yun, L. (2021). Method of reaching consensus on probability of food safety based
on the integration of finite credible data on block chain. IEEE Access, 9, 123764–123776.
Yang, C.-Z., Ma, J., Wang, S.-L., & Liew, A. W.-C. (2020). Preventing deepFake attacks on speaker authentication by dynamic lip movement
analysis. IEEE Transactions on Information Forensics and Security, 16, 1841–1854.
Yang, D., Zhu, T., Wang, S., Wang, S., & Xiong, Z. (2022). LFRSNet: A robust light field semantic segmentation network combining contex-
tual and geometric features. Frontiers in Environmental Science, 10, 996513.
Yang, J., Xiao, S., Li, A., Lan, G., & Wang, H. (2021). Detecting fake images by identifying potential texture difference. Future Generation
Computer Systems, 125, 127–135.
Yazdinejad, A., Parizi, R. M., Srivastava, G., & Dehghantanha, A. (2020). Making sense of blockchain for AI deepfakes technology. Paper
presented at the 2020 IEEE Globecom Workshops (GC Wkshps).
Yu, I.-J., Nam, S.-H., Ahn, W., Kwon, M.-J., & Lee, H.-K. (2020). Manipulation classification for jpeg images using multi-domain features.
IEEE Access, 8, 210837–210854.
Yu, P., Xia, Z., Fei, J., & Lu, Y. (2021). A survey on Deepfake video detection. IET Biometrics, 10(6) 607–624.
Zadeh, F. A., Bokov, D. O., Yasin, G., Vahdat, S., & Abbasalizad-Farhangi, M. (2023). Central obesity accelerates leukocyte telomere length
(LTL) shortening in apparently healthy adults: A systematic review and meta-analysis. Critical Reviews in Food Science and Nutrition, 63
(14), 2119–2128.
Zhang, M., Chen, Y., & Lin, J. (2021). A privacy-preserving optimization of neighborhood-based recommendation for medical-aided diagno-
sis and treatment. IEEE Internet of Things Journal, 8(13), 10830–10842.
Zhang, M., Chen, Y., & Susilo, W. (2020). PPO-CPQ: A privacy-preserving optimization of clinical pathway query for e-healthcare systems.
IEEE Internet of Things Journal, 7(10), 10660–10672.
Zhang, W., Zhao, C., & Li, Y. (2020). A novel counterfeit feature extraction technique for exposing face-swap images based on deep learning
and error level analysis. Entropy, 22(2), 249.
Zhang, Y., Gao, F., Zhou, Z., & Guo, H. (2021). A survey on face forgery detection of Deepfake. Paper presented at the Thirteenth International
Conference on Digital Image Processing (ICDIP 2021).
Zhang, Z., Luo, C., & Zhao, Z. (2020). Application of probabilistic method in maximum tsunami height prediction considering stochastic sea-
bed topography. Natural Hazards, 104(3), 2511–2530.
Zhao, H., Zhou, W., Chen, D., Wei, T., Zhang, W., & Yu, N. (2021). Multi-attentional deepfake detection. Paper presented at the Proceedings
of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.
Zhao, Z., Wang, P., & Lu, W. (2021). Multi-layer fusion neural network for Deepfake detection. International Journal of Digital Crime and
Forensics (IJDCF), 13(4), 26–39.
Zheng, W., Liu, X., Ni, X., Yin, L., & Yang, B. (2021). Improving visual reasoning through semantic representation. IEEE Access, 9, 91476–
91486.
Zheng, W., Liu, X., & Yin, L. (2021). Sentence representation method based on multi-layer semantic network. Applied Sciences, 11(3), 1316.
Zheng, W., Xun, Y., Wu, X., Deng, Z., Chen, X., & Sui, Y. (2021). A comparative study of class rebalancing methods for security bug report
classification. IEEE Transactions on Reliability, 70(4), 1658–1670.
Zheng, W., Yin, L., Chen, X., Ma, Z., Liu, S., & Yang, B. (2021). Knowledge base graph embedding module design for visual question answer-
ing model. Pattern Recognition, 120, 108153.
Zhong, L., Fang, Z., Liu, F., Yuan, B., Zhang, G., & Lu, J. (2021). Bridging the theoretical bound and deep algorithms for open set domain
adaptation. IEEE Transactions on Neural Networks and Learning Systems.
Zhou, L., Fan, Q., Huang, X., & Liu, Y. (2023). Weak and strong convergence analysis of Elman neural networks via weight decay regulariza-
tion. Optimization, 72(9), 2287–2309. https://doi.org/10.1080/02331934.2022.2057852
Zhou, X., Wang, Y., & Wu, P. (2020). Detecting deepfake videos via frame serialization learning. Paper presented at the 2020 IEEE 3rd
International Conference of Safe Production and Informatization (IICSPI).
Zi, B., Chang, M., Chen, J., Ma, X., & Jiang, Y.-G. (2020). Wilddeepfake: A challenging real-world dataset for deepfake detection. Paper
presented at the Proceedings of the 28th ACM International Conference on Multimedia.
Zong, C., & Wan, Z. (2022). Container ship cell guide accuracy check technology based on improved 3D point cloud instance segmentation.
Brodogradnja: Teorija i Praksa Brodogradnje i Pomorske Tehnike, 73(1), 23–35.
Zong, C., & Wang, H. (2022). An improved 3D point cloud instance segmentation method for overhead catenary height detection.
Computers & Electrical Engineering, 98, 107685.

How to cite this article: Heidari, A., Jafari Navimipour, N., Dag, H., & Unal, M. (2024). Deepfake detection
using deep learning methods: A systematic and comprehensive review. WIREs Data Mining and Knowledge
Discovery, 14(2), e1520. https://doi.org/10.1002/widm.1520