Journal of Cyber Security Technology

ISSN: (Print) (Online) Journal homepage: www.tandfonline.com/journals/tsec20

A systematic literature review on the effectiveness

of deepfake detection techniques

Laura Stroebel, Mark Llewellyn, Tricia Hartley, Tsui Shan Ip & Mohiuddin
Ahmed

To cite this article: Laura Stroebel, Mark Llewellyn, Tricia Hartley, Tsui Shan Ip &
Mohiuddin Ahmed (2023) A systematic literature review on the effectiveness of
deepfake detection techniques, Journal of Cyber Security Technology, 7:2, 83-113, DOI:
10.1080/23742917.2023.2192888

To link to this article: https://doi.org/10.1080/23742917.2023.2192888

© 2023 The Author(s). Published by Informa

UK Limited, trading as Taylor & Francis
Group.

Published online: 29 Mar 2023.

Submit your article to this journal

Article views: 3950

View related articles

View Crossmark data

Citing articles: 17 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tsec20
JOURNAL OF CYBER SECURITY TECHNOLOGY
2023, VOL. 7, NO. 2, 83–113
https://doi.org/10.1080/23742917.2023.2192888

REVIEW ARTICLE

A systematic literature review on the effectiveness of

deepfake detection techniques
Laura Stroebel, Mark Llewellyn, Tricia Hartley, Tsui Shan Ip
and Mohiuddin Ahmed
School of Science, Edith Cowan University, Joondalup, Western Australia, Australia

ABSTRACT ARTICLE HISTORY

With technological advances, the generation of deepfake mate­ Received 29 November 2022
rial is now within reach of those operating consumer-grade Accepted 14 March 2023
hardware. As a result, much research has been undertaken on KEYWORDS
deepfake detection techniques. This work has analysed and Deepfake detection
measured the performance of various detection techniques techniques; machine
using multiple metrics and discussed the effectiveness of learning; deep learning;
these deepfake detection techniques. This has been undertaken methods; dataset
by examining and analysing the current state of deepfake
detection techniques. Unlike other existing surveys, this work
produced a Systematic Literature Review (SLR) on research
conducted from the beginning of 2021 to August 2022. This
SLR includes tabulated data containing details of the techniques
used and the accuracy of those techniques, performance
metrics provided in each study, summaries of the datasets
used, and challenges and future trends. This SLR has been
undertaken with a focus on using a mixed methods approach.
This SLR has determined that deep learning (DL) has surpassed
machine learning (ML) as the preferred deepfake detection
model. However, ML is still a primary focus method in medical
imagery. It was also discovered that traditional artificial neural
networks are no longer effective and require additional mod­
ules to produce ensembled and multi-attentional architectures.

1. Introduction
The remarkable technological advancements, especially in artificial neural net­
works, coupled with high computing power, have led to the creation of tech­
nologies that can be used to tamper with digital content. These technologies,
including FakeApp and FaceApp, have been used intensely in the recent past to
generate real-looking, but fake, visual content. Most of these technologies
enable a creator to alter various attributes in image and video content, like
hairstyle, age, voice, and many others, or even to swap the entire face in the

CONTACT Mohiuddin Ahmed [email protected] School of Science, Edith Cowan University Perth,
Western, Australia
© 2023 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group.
This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivatives
License (http://creativecommons.org/licenses/by-nc-nd/4.0/), which permits non-commercial re-use, distribution, and reproduc­
tion in any medium, provided the original work is properly cited, and is not altered, transformed, or built upon in any way. The
terms on which this article has been published allow the posting of the Accepted Manuscript in a repository by the author(s) or
with their consent.
84 L. STROEBEL ET AL.

image with another. This has led to the concept of Deepfake, a name coined
from Deep Learning and Fake, which exploits the powers of deep learning to
create fake, realistic video and image content. Most of these deepfake technol­
ogies result from two core neural network technologies: generative network
and discriminative network which when combined, give rise to Generative
Adversarial Networks (GANs).
Deepfake content has a wide range of useful applications across many
industries, that are yet to be fully harnessed. Abdulreda and Obaid (2022),
in their study [2], discuss techniques of facial modification. They argue
that Entire Face Synthesis is a facial modification technique used to
produce non-existent faces. Methods such as StyleGAN utilise GANs and
can produce high-quality face pictures that can be utilised in video games
and 3D modelling. However, the technology has been widely used by
malicious actors for malicious intent. For example, as early as 2017,
deepfake technologies were used to share pornographic videos that had
celebrity faces as actors. Also, these technologies have been used to
spread misinformation in various avenues, mainly across social media
platforms, conducting cybercrimes and creating political tension and
instability.
This has prompted the need to detect and distinguish between what is real
and what is fake. So far, much research has been done on deepfake detection
techniques and the ability of these techniques to identify deepfake images and
videos using artificial intelligence (AI) models trained on well-established data­
sets of deepfake and typical examples. However, there has been limited
research undertaken on the validity of the training datasets used by these AI
models which, due to a lack of annotation of the features and potential bias of
the content, may cause the deepfake detection tool to fail [3].
Bias in the datasets used for AI and machine learning has been suspected for
some time and studies suggest AI models trained on unbalanced datasets will be
biased against certain groups and will perform poorly. Deepfake detection tech­
nologies using these AI models will sometimes fail to identify deepfake content due
to the lack of diversity in the dataset they are trained on. Further to this, Xu et al.
(2022) propose that the deepfake detection tools themselves have a bias that
influences results and this is further exacerbated by the bias in the dataset used [4].
The limited number of audio datasets, particularly non-English based, is
another area where detection of deepfake media is lacking. Adding to this,
with the impact of accents in the current audio datasets yet to be investigated,
and audio deepfakes increasing, the need for further study in this area is
paramount.
Table 1 summarises the literature reviews/surveys that have already been
undertaken on this subject – these will be discussed in more detail in section 3.1.
The main contributions of this SLR are:
 JOURNAL OF CYBER SECURITY TECHNOLOGY 85

● A comprehensive survey of current literature relating to deepfake detec­

tion technologies – specifically, those produced between January 2021 and
August 2022.
● A comprehensive analysis and discussion of new deepfake detection tech­
niques. This study details the most recent techniques and provides perfor­
mance metrics for them.
● Discussion of the challenges faced in deepfake detection techniques.
● Gives highlights of what the future of deepfake detection will look like
based on the most recent advancement in the domain.

This paper has been organised in the following manner:

● Section 2: The systematic literature review (SLR) process, including the

criteria used to identify viable source documentation and the research
questions asked.
● Section 3: Discuss other SLR’s undertaken on this subject and overview
modern deepfake detection methods.
● Section 4: Discuss this SLR’s findings and introduce some limitations,
challenges and future trends identified.
● Section 5: Conclusion.

2. Systematic literature review process

This SLR aimed to create a body of knowledge of deepfake detection techniques
and to conduct a systematic review of the currently available literature regard­
ing these techniques. The main objective was to undertake an SLR that analyses
the effectiveness of deepfake detection techniques [1]. Our specific objectives
were to:

● Examine and analyse the current state of deepfake, providing an up-to-

date overview of recent research work on deepfake detection.
● Analyse and measure the performance of various deepfake detection
techniques using multiple metrics; and
● Identify and discuss significant advances, challenges, and future trends.

To achieve these objectives, it was important to conduct our study in an

organised and systematic way. The section below illustrates our approach.

2.1. Search strategy

An important part of this literature review was to gather existing publica­
tions that proposed deepfake detection techniques. We did not cover
deepfake generation techniques. The initial search strategy was based
 86

Table 1. Summary of previous works in comparison to our study. (ref)* means references listed as reviewed document numbers are not noted in the paper and thus
the number is inferred from a paper’s reference section. The (-) under the generation and detection models means the study did not explore the technology under
that column.
L. STROEBEL ET AL.

Prior Literature Deepfake Documents Generation Detection Datasets

Reviews/Surveys Researched Completed Date Reviewed Models Models Identified Challenges
Weerawardana and Detection August 2021 48 (ref)* - 17 7 No definitive deepfake detection method identified and lack of quality
Fernando [5] datasets
Zhang [6] Generation/ September 2021 99 (ref)* 12 25 15 Lack of benchmark test methods and quality datasets
Detection
Malik et al. [7] Generation/ January 2022 130 (ref)* 17 40 17 Include lack of datasets, unknown types of attacks, temporal aggregation
Detection and unlabelled data
Juefei-Xu et al. [8] Generation March 2022 318+ 91 117 24 Lack of quality datasets, lack of competitive baselines, generalisation and
Detection robustness, evaluation platform and metrics
2016 to 2021
Rana et al. [9] Detection February 2022 112 - 91 18 Need a framework to reduce inconsistencies and to produce definitive and
2018 to 2020 systematic outcomes
Celebi et al. [10] Detection February 2022 20 (ref)* - 9 7 There are 100 times more people working on deepfake generation than on
detection
Almutairi and Detection May 2022 57 (ref)* - 22 8 Limited non-English detection methods, lack of research on accents and
Elgibreen “noise”, excessive pre-processing required
(2022) [11]
Masood et al. (2022) Generation/ May 2022 436 100 67 16 Generalisation,
[12] Detection Temporal coherence, Datasets integrity,
Up to Temporal aggregation
March 2022
 JOURNAL OF CYBER SECURITY TECHNOLOGY 87

upon the strategy presented in [9] and was conducted using the standard
research databases. Table 2 shows the search strategy used in this paper.
Further investigation on identified areas of interest was done via the data
sources in Table 2 as well as other scholarly literature databases such as arXiv
and Research Gate. Figure 1 shows the search results for this SLR – made up
of 83 documents reviewed and considered relevant for this SLR, 21 artifacts
collected for reference purposes only and 16 documents reviewed and found
to not fit the scope of this SLR based on the documents creation date or
content.
Figure 2 shows the source type of the articles included in the Reviewed Works
section of this document.

2.2. Research questions

We developed research questions to be used to assess the current literature and

to gather relevant information from these sources. These questions also served
the purpose of focussing our review on the areas we believed to be most
relevant and important to this area of research. The right questions needed to
be asked, so that the research community could benefit from useful, targeted
information. Table 3 outlines the research questions, as well as their purpose.

2.3. Data extraction and recording

Due to the large amount of data to collect, sort through and categorise, it was
important to do this in an organised and meaningful way. This enabled us to
monitor the data, ensure information was not missed and be able to extract
meaning from the data.

Table 2. SLR search strategy.

Type Description
Aim Gather existing literature that outlined deepfake detection techniques, not limited to video
and audio.
Data Sources The majority of articles were located in: IEEE Publications Database, ProQuest Central,
ScienceDirect and SpringerLink.
Query A keyword search was conducted using the following query: deepfake OR faceswap OR
“video manipulation” OR “fake face” OR “fake image” OR “fake video” AND detection OR
detect OR “facial manipulation” OR “digital media forensics”. Articles were limited to
those published in 2021 and 2022.
NB: A search of titles only was attempted however there were too few results.
Method Using the search query above, we obtained 411 search results. The abstracts were reviewed
to determine if the publication related to deepfake detection techniques. These were
then entered into a spreadsheet after which they were reviewed in more detail.
Inclusions/ Books were excluded, as were articles that did not contain information on deepfake
exclusions detection or metrics for the deepfake detection technique discussed. Also, excluded were
documents published before 2021.
88 L. STROEBEL ET AL.

Figure 1. Documents returned by our search strategy and how they were categorised after
review.

Figure 2. The source type of the documents (by year) included as reviewed works.
 JOURNAL OF CYBER SECURITY TECHNOLOGY 89

Table 3. Research questions.

Number Question Purpose
RQ1 What are the current deepfake detection Determine the currently proposed detection
techniques? techniques.
RQ1A What is the effectiveness of these Assess the effectiveness of the detection techniques
techniques? using common metrics.
RQ2 What advances have occurred since the Describe the positive changes that have occurred.
previous SLRs?
RQ3 What are the current challenges involved in Identify areas that challenge successful deepfake
deepfake detection? detection.
RQ4 Can any future trends be identified, and if Identify trends in deepfakes and their detection.
so, what are they?

2.4. Data analysis and visualisation

The purpose of this phase was to review, compare and analyse the data
obtained from all sources. This allowed us to generate insights and determine
any trends that were present. The data was presented visually using methods
such as pie charts and tables.

3. Reviewed works
For this SLR the works reviewed could be categorised in two ways – previous
systematic literature reviews/surveys and modern deepfake detection
methods.

3.1. Previous literature reviews/surveys

We reviewed 8 literature review/survey documents − 1 published in 2021 and 7
published in 2022. The majority included deepfake generation and detection
techniques [6–8,12] and the remaining focused only on deepfake detection
techniques [5,9–11]. However, while Celebi et al. (2022) discussed deepfake
detection techniques, it was in the context of evidence for judicial proceedings,
rather than specific models [10].
Masood et al. (2022) surveyed publications relating to audio/video manipula­
tion, specifically generation and detection techniques available up to
March 2022. The comprehensive study [12] covered works from 2017 to 2022,
resulting in a total of 436 papers being reviewed. It [12] found some of the
limitations of generation processes were generalisation across source datasets,
pose variations, distance from the camera scenarios, and temporal coherence.
While the challenges identified with deepfake detection technologies were the
quality, fairness, and trust of deepfake datasets, temporal aggregation, and
social media laundering, - to name a few.
Juefei-Xu et al. (2022) focused solely on facial deepfakes, thoroughly examin­
ing available generation and detection techniques, relating to this area, up to
June 2021. This paper [8] noted that deepfake detection techniques were
90 L. STROEBEL ET AL.

proportionate to deepfake generation techniques and vice versa – that is, the
creation of generation techniques was driven as much by the creation of
detection techniques as the other way around (detection by generation). This
survey [8] noted a need for competitive baselines to evaluate deepfake detec­
tion techniques so true performance could be assessed. Zhang (2022) agreed,
indicating the deepfake detection techniques up to 2020, reviewed in this
survey [6], were not robust and there were many challenges when assessing
generation and detection models, including a lack of benchmark test methods
and quality datasets.
Malik et al. (2022) took a similar approach, evaluating face image and video
deepfake techniques (generation and detection) up to early 2021, and con­
cluded there was a general inability in the detection models to transfer and
generalise indicating further research was needed [7].
Rana et al. (2022) undertook a detailed SLR of 112 articles relating to deep­
fake detection technologies, published between 2018 and 2020, inclusive. They
[9] classified these technologies into 4 categories (deep learning-based techni­
ques, classical machine learning-based methods, statistical techniques, and
blockchain-based techniques) and evaluated the performances of each when
used with the datasets available at the time. At the conclusion of this SLR [9],
deep learning-based methods were identified as outperforming all other cate­
gories. This conclusion was also noted by Weerawardana and Fernando (2021),
who categorised the deepfake detection techniques they reviewed into tradi­
tional and deep learning methods only [5].
Almutairi and Elgibreen (2022) identified three types of audio deepfakes –
synthetic-based, imitation-based and replay-based. This survey [11] examined
machine and deep learning audio deepfake detection technologies and deter­
mined that machine learning methods proved more accurate than deep learn­
ing, but required excessive training and manual feature extraction, potentially
making them unscalable. It [11] noted a lot more study in this area was needed
to address existing gaps and challenges.

3.2. Modern deepfake detection methods

We reviewed 75 modern deepfake detection technology documents to assess

developing trends and identify future research areas. Of these, the majority
focused on video, image, or a combination of image/video deepfakes (90%). The
remaining documents looked at audio or a combination of audio/video and all
media.
There were several machine learning models proposed using a generative
adversarial network (GAN) framework [13–15]. However, the majority were deep
learning architectures based on artificial neural networks – such as convolu­
tional neural network (CNN), deep neural network (DNN) and long short-term
memory (LSTM). Table 4 details the deepfake detection techniques discussed in
 JOURNAL OF CYBER SECURITY TECHNOLOGY 91

the reviewed documents. In total, 48 deepfake detection models were identified

that met the objectives of this SLR. Figure 3 is a cluster map of the noteworthy
deepfake detection techniques and shows that most models were CNN based.
An overview of the identified techniques will be delivered in the following
structure:

● Machine learning deepfake detection techniques.

● CNN, DNN and LSTM techniques used as standalone models and those
extended to multifaceted architectures.
● Techniques to detect audio deepfakes.
● Datasets used to train, test and validate deepfake detection models -
Table 6 details the datasets used by the deepfake detection models
reviewed in this SLR.

3.2.1. Machine learning

MSTA_Net, proposed by Yang et al. (2022) was a machine-learning deepfake
detection model which examined the texture features of an image to identify
anomalies. This model [13] used the whole image in the detection process, not
just focused facial areas, to link the forged and non-forged areas of the image.
That is, if inconsistency in the image texture was found the image was tagged
fake, but if no inconsistency was found, the image was tagged as non-fake.
Other machine learning models included a self-supervising decoupling net­
work model suggested by Zhang et al. (2021) that utilised dual feature learning
and performed well with lower-quality images [86]. As well as a hybrid texture/
noise model proposed by Fu et al. (2022) which used local binary pattern and
subtractive pixel adjacency matrix features to detect and identify facial manip­
ulation in deepfake images [17].

3.2.2. CNN, DNN, LSTM and RNN

EfficientNet-V2, discussed by Deng et al. (2022) was a CNN model evolved from
the EfficientNet series first introduced in 2019. This more streamlined version
[44] used fewer resources to produce higher test accuracy. It also could be used
to identify multiple faces in an image and track them across multiple frames.
However, it was noted [44] that the pre-processing required by the model
caused some inconsistencies which needed manual intervention, and the
model was not tested for generalisability. Wang et al. (2021) used EfficientNet
as the testing model for the adversarial training framework they proposed [43].
Other CNN models discussed, were various versions of XceptionNet
[18,22,47,87] and ResNet [20,21,23,49]. A unique 3-layer frequency CNN was
proposed by Kohli et al. (2021) to work with a two-dimensional global discrete
Cosine transform model [27]. As well as a lightweight 3D CNN [24] model which
extracted features in the spatial-temporal dimension, proposed by Liu et al.
(2021).
Table 4. The deepfake detection techniques proposed in the reviewed documents.
92

Metrics (NB: averaged if

DFD DF Media Source Year multiple results)
Techniques DFD Models/Methods DFD Dataset Researched Documents Pub DF Type AUC (%) ACC (%)
Machine Learning Models
SSDN (Self-supervised decoupling network) FF++ Image [16] 2021 FM 98 (HQ)
91.8 (LQ)
GAN MSTA-Net – multi-scale self-texture attention generative network FF++, CDF, Image [13] 2022 FS FF++ 92 (HQ)
DFDC, DFor-1.0 FF++ 88 (LQ)
SVM Hybrid Texture/Noise, SVM – using local binary pattern and subtractive FFHQ Image [17] 2022 FM 99.4 97.6
L. STROEBEL ET AL.

pixel adjacency matrix features

Deep Learning Models
CNN XceptionNet – comparison test FF++ Video [18] 2021 FM FF++ 99.3
CNN Yolo (V2) SYSU-OBJFORG, SULFA Video [19] 2021 VM 99
CNN ResNet-50 - face recognition model pretrained on MS1M-ArcFace CDF, FF++ [20] Image [20,21] 2021 FS CDF 98 [20] 97 [21]
/WebFace12M [20] RFFD [21] FF++ 99
Comparison test [21] [20]
99.7 [21]
CNN Passive/Proactive method FF++ Video [22] 2021 FM highest result 90
(Bilateral filtering/Joint adversarial training – XceptionNet/ (over many
MesInceptionNet) techniques/
metrics)
CNN Three-part architecture – using separate image noise analysis on FF++, CDF Image [23] 2021 FM FF++ 99 FF++ 99
manipulated and blending artifacts in an image CDF 99 CDF 99
FF++/CDF FF++/CDF 81
81
CNN 3D CNNs FF++, DFTIMIT, DFDC, Video [24,25] 2021 FM FF++ 99 FF++ 99/94 (HQ/
CDF [24] [24] LQ) [25]
FF++, VidTIMIT [25] DFTIMIT 99 VidTIMIT 99/99
[24] (HQ/LQ) [25]
DFDC 93
[24]
CDF 98 [24]
CNN Shallow-FakeFaceNet (SFFN) Own dataset HFM, Image [26] 2021 FM 64×64 94
(Shallow CNN) RFFD, CelebA 1024×1024
99.9
CNN fCNN with two-dimensional global discrete Cosine transform (2D-GDCT) FF++, CDF Video [27] 2021 FM CDF 75.2 CDF 66.5
FF++ 96.7 FF++ 89.28
DNN Dual path system – using transformer and node-compressed DFDC, CDF, FF++ Video [28] 2021 FM DFDC 94.9 FF++ 95.5
architecture CDF 92.3 DFDC 88.4
CDF 74.3
(Continued)
Table 4. (Continued).
Metrics (NB: averaged if
DFD DF Media Source Year multiple results)
Techniques DFD Models/Methods DFD Dataset Researched Documents Pub DF Type AUC (%) ACC (%)
CNN++ Bita-Net- utilising ResNet-50 and U-Net Trained on DFDC Image/ [29] 2021 FM FF++ 99
FF++, CDF, Video CDF 99
DeepfakeTIMIT, UADFV DFTIMIT 99
UADFV 99
CNN++ FST-Net (dual-stream architecture) - utilising ResNet-152 and S3D Trained on ImageNet Video [30] 2021 FM 97 (LQ) 99 (RAW)
(separable 3D CNN) FF++ 98 (HQ)
93 (LQ)
CNN++ AMTENnet – a combination of AMTEN and CNN HFF, FF++ Video [31] 2021 FS HFF 99
CNN++ FeatureTransfer – two-stage adversarial training pipeline Trained on ImageNet Video [32] 2021 FM (averaged
FF++, DFTIMIT, DFDC, across
CDF,DFD many
products)
FF++ 93
DFTIMIT 96
DFDC 79
CDF 98
DFD 89
CNN++ AFIFN (Advanced Fake Image-Feature Network) using DCT-based CelebA Image [33] 2021 FM 96.1
processing and Y Cr Cb based pre-processing
CNN++ CWSA-Net (Channel-Wise Spatiotemporal Aggregation) with FF++, CDF, DFDC Video [18] 2021 FM FF++ 99.6 FF++ 99.4
EfficientNet- B0 backbone CDF 99.7 CDF 95.9
DFDC 92.5 DFDC 83.7
CNN++ PGT, SWDCT_IDCT FF++, CDF, WDF Image, [34] 2021 FS FF++ 1.0 FF++ 99.8
video CDF 99.7 WDF 79.8
WDF 87.8
CNN++ LRNet UADFV, FF++, CDF, Video [35] 2021 VM FF++ 99.9 Dfor-1.0 97.74
Dfor-1.0 UADFV 98.5 FF++ 99.7
CDF 56.9
CNN++ DCWNet (dual-tree complex wavelet-based face forgery network) with FF++ Image [36] 2021 FM 99 (HQ) 98.7 (HQ)
JOURNAL OF CYBER SECURITY TECHNOLOGY

ResNet-34 backbone 99 (LQ) 97.9 (LQ)

CNN++ DeepfakeNet – combination of ResNet and Inception model FF++, Kaggle, DFTIMIT Video [37] 2021 FM FF++ 96 FF++ 96.7
(Continued)
93
 94

Table 4. (Continued).
Metrics (NB: averaged if
DFD DF Media Source Year multiple results)
Techniques DFD Models/Methods DFD Dataset Researched Documents Pub DF Type AUC (%) ACC (%)
CNN++ Multi-dimensional biological signals FF++, DFD, UADFV Video [38] 2021 FM FF++ 96
DFD 97
UADFV 95
L. STROEBEL ET AL.

CNN++ SiteForge CASIA 2.0 Image [39] 2021 FM/IM CASIA 94.7
CNN with added Local Interpretable Model-agnostic Explanations Own dataset Twitter Own dataset 83.2
(LIME) Indian Dataset 2.0
CNN++ ResNet-18 - pre-processed with Image saliency and guided filter FF++ Image [40] 2021 FM FF++ 99
processing
SCNN Set Convolutional Neural Network (SCNN) using MesoNet/XceptionNet DFTIMIT, FF++, DFDC-P Video [41] 2021 FM 80/99 (HQ) 80/96(HQ)
79/95 (LQ) 80/94 (LQ)
CNN TCN – temporal convolution network FoR [42] Audio [11,42] 2021 FM 92 [42]
2022
CNN STN – spatial transformer network FoR [42] Audio [11,42] 2021 FM 80 [42]
2022
CNN EfficientNet-B0 - adversarial trained on FF++ FF++ Image/ [18,43] 2022 FM 69 [43]
Comparison test DFDC training Video 2021 96 (many datasets)
[18]
CNN EfficientNet-V2 FF++, FFIW-10K Video [44] 2022 FS FF++ 98
FFIW-10K 93
DNN MC-LCR – multimodal contrastive classification by locally correlated FF++, CDF, DFDC, DFor- Video [45] 2022 FM FF++ 99/90 FF++ 98/88
representations 1.0 FF++/DFDC FF++/CDF 71
Xception (pre-trained on ImageNet), MLP-Mixer 71 FF++/DFDC 70
FF++DFor-1.0 75
DNN Face/background squares captured/Siamese noise trace extraction/ CDF, UADFV Video [46] 2022 FS CDF 99.92 99.15
noise similarity analysis UADFV
88.95
CNN + ViT ViXNet FF++, CDF, DFID Image [47] 2022 FM FF++ 99/75 FF++ 97/69
CDF 99/73 CDF 94/67
DFID 99/75 DFID 95/68
(Continued)
Table 4. (Continued).
Metrics (NB: averaged if
DFD DF Media Source Year multiple results)
Techniques DFD Models/Methods DFD Dataset Researched Documents Pub DF Type AUC (%) ACC (%)
CNN++ ID-Reveal − 3D morphable model (3DMM) + Temporal ID Network + DFD Video [48] 2022 FM 87/96 (FR/ 76/85 (FR/FS)(HQ)
GAN FS)(HQ) 82/78 (FR/FS)(LQ)
90/94 (FR/
FS)(LQ)
CNN++ Dual-level collaborative framework (un-named) - utilising ResNet-50 Trained on ImageNet Video [49] 2022 FM FF++ 99 FF++ 95
and temporal learning models FF++, CDF, DFDC CDF 98 CDF 96
DFDC 84 DFDC 99
CNN++ DenseNetXX – extends ResNet by adding transition layer RFFD Image [21,50] 2022 FM [50] 99 [50] 94 [50]
FS [21] DN121 97.1 DN121 97 [21]
[21] DN169 95 [21]
DN169 99.6 DN201 96 [21]
[21]
DN201 99.4
[21]
CNN++ VGG-XX – fully connected classifiers followed by maxpooling layers RFFD Image [21,50] 2022 FM [50] VGG-19 96 VGG-19 95
FS [21] [50] VGG-19 94 [21]
VGG-19 VGG-16 92 [21]
98.7 [21] VGG-Face 99 [21]
VGG-16
97.7 [21]
VGG-Face
99.8 [21]
CNN++ CustomCNN – analyses dropout, padding, augmentation and grayscale RFFD Image [50] 2022 FM 98 89
on model performance
CNN++ LiSiam (Localisation Invariance Siamese Networks) using Xception FF++, CDF Video [51] 2022 FM 99 (HQ) 96 (HQ)
backbone 91 (LQ) 87 (LQ)
FF++/CDF
81
CNN++ Multilayer Perceptron Video [52] 2022 FM 87 87
JOURNAL OF CYBER SECURITY TECHNOLOGY

CNN++ Frame-temporality two-stream convolutional network – with CDF, FF++ Video [53] 2022 FS CDF 87 CDF 80.74
MesoNet – for frame-level and ResNet-18 for time-dependent FF++ 94 FF++ 86.61
residual features
(Continued)
95
 96

Table 4. (Continued).
Metrics (NB: averaged if
DFD DF Media Source Year multiple results)
L. STROEBEL ET AL.

Techniques DFD Models/Methods DFD Dataset Researched Documents Pub DF Type AUC (%) ACC (%)
CNN++ Patch-DFD – using RESNet-50 and Inception-v3, pre-trained on DFTIMIT, CDF, FF++ Image [54] 2022 FM DFTIMIT FF++ 96.23/87.36
ImageNet 99.42/99.1
CDF 98.88
RNN Xception-ConvLTSM – with spatiotemporal attention mechanism FF++, CDF, DFDC Video [55] 2022 FS FF++ 99 FF++ 99
(CNN CDF 99 CDF 99
+LSTM) DFDC 94 DFDC 92
RNN EfficientNet-B3 [56] FF++ [56] Video [56] [56,57] 2022 FS Celeb-DF 99 FF++ 99 (raw) [56]
(CNN With vanilla LSTM [57] Celeb-DF [57] Image [56,57] [57] FF++ 99 (HQ) [56]
+LSTM) [57] FM [56] FF++ 91 (LQ) [56]
Celeb-DF 93.9 [57]
RNN Meso-4 with vanilla LSTM Celeb-DF Image [57] 2022 FM Celeb-DF 96 Celeb-DF 89.3
(CNN
+LSTM)
RNN ResNet and LSTM CDF Video [58] 2022 FM CDF 88.8 CDF 91
(CNN
+LSTM)
Other Models
Jointly modelling video and audio modalities for deepfake detection FF++, DFDC Video and [59] 2021 AM/ FF++ FF++ 95
Audio VM 99
PSCC-Net [60] Progressive Spatio-Channel Correlation Network Own dataset Image [61] 2022 AM 99.6
EfficientNet and Vision Transformers [62] FF++, DFDC Video [62] 2022 VM 95.1
 JOURNAL OF CYBER SECURITY TECHNOLOGY 97

Figure 3. Cluster map of identified deepfake detection models.

Many papers combined the traditional techniques (CNN, DNN, LTSM) with
other modules to produce ensembled and multi-attentional architectures. For
example, we found a number of models that extended various versions of
ResNetXX. Bita-Net, created by Ru et al. (2021), was based on human detection
methods – using a temporal examination at high frame rate and frame by frame
scrutiny of key frames with an additional attention branch- using ResNet-50 and
U-Net. This model [29] produced excellent, consistent results across multiple
datasets, even when trained by one dataset and then validated and tested on
another.
Pu et al. (2022) examined videos with frame-level and video-level methods
to identify fake content. ResNet-50 extracted facial features, and then fed
them to a temporal learning model and a video and frame level classifier, in
a three-step examination process to determine if a video was fake or not.
This model [49] was touted to outperform all existing methods for frame and
video level detection, as well as being robust to video quality and database
variations. However, it [49] was not generalisable to other types of facially
manipulated images or videos, such as those that were GAN-generated
(completely made-up).
98 L. STROEBEL ET AL.

Table 5. The description of the acronyms used in Table 4.

Acronym Meaning
AM Audio Manipulation
CNN + ViT Convolution neural network extended with a vision transformer
CNN++ Convolution neural network extended with other modules/components
FM Facial Manipulation
FS Face Swapping
IM Image Manipulation
LSTM++ Long short-term memory extended with other modules/components
SCNN Set convolutional neural network
SVM Support vector machine
VM Video Manipulation

Zhang et al. (2021) proposed a dual-stream architecture, called FST-Net,

made up of frequency and spatial-temporal streams, enhanced with an
attention mechanism on each, and using ResNet-152. This model [30] proved
successful across high and low-quality videos but had limitations when used
with some datasets, which reduced its generalisability. Future work [30]
suggested for this model involved incorporating synthetic audio detection
to improve accuracy. Along a similar line, Chen et al. (2022) proposed
a spatiotemporal attention mechanism that worked with an LSTM to enhance
its functionality. This model [55] had a 4 step architecture made up of
a spatiotemporal attention mechanism, a CNN, an LSTM and a predictor to
make the final decision.
Other models which used ResNetXX as their backbone were DenseNetXX –
a model used for comparative testing in several papers [21,50]. A two-stream
model created by Hu et al. (2022) analysed frame-temporality streams in videos
and using ResNet-18 [53]. A dual-tree complex wavelet transform-based net­
work [36] with ResNet-34 as the backbone of the two streams was proposed by
Gao et al. (2021). DeepfakeNet, based on a combination of ResNet’s stacking
feature and Inception’s split-transform-merge feature [37], was introduced by
Gong et al. (2021). This technique was also discussed by Yu et al. (2022) using
ResNet-50 and Inception-v3 [54].
ID-Reveal created by Cozzolino et al. (2022), combined a facial feature
extractor, a temporal network for biometric anomaly detection and a GAN to
predict expression-based motion. This model [48] successfully identified deep­
fakes in high and low-quality videos, producing good results in detecting face-
swapping instances but less so for other types of facial manipulation. This model
[48] needed to be trained on pristine videos of the source identity, which was
limiting, but meant it was generalisable and robust when used with other
datasets.
MC-LCR, created by Wang et al. (2022) was a DNN using a multimodal
contrastive classification by locally correlated representations [45]. It was a two-
stream framework working in the spatial and frequency domains that, while
 JOURNAL OF CYBER SECURITY TECHNOLOGY 99

giving good results on high and low-quality images, seemed to struggle with
varying light conditions and biometric challenges.
A dual-path system proposed by Luo et al, (2021) achieved good results using
neural ordinary differential equations (NODE) and a facial feature transformer
architecture [28]. However, the technique was suitable for facial manipulation
deepfakes only as it focused too much on local facial features meaning it did not
perform well on face-swapping deepfakes.
Vamsi et al. (2022) created an RNN model called ConvNets, consisting of
ResNet and LSTM [58]. The paper [58] noted, due to the architecture of the
model, less pre-processing was required and images were handled in a way that
made the model fast to implement and easy to use. It [58] also claimed high
performance when compared to other CNN models using the same datasets.
Xu et al. (2021) through extensive experimentation with multiple variations of
detection techniques developed an SCNN which gave robust and verifiable
results. The authors asserted that the detection technique [41] was more robust
than previous models and potentially could be integrated with other backbone
networks to enhance performance.
Raskar et al. (2021) proposed YOLO which claimed to be extremely accurate
and efficient in terms of computational time in object-based video forgery
detection. This technique [19] detected complex copy-move video deepfakes
with 99% accuracy. This technique was best suited to copy-move attacks
including scaling, rotation and flipping [19]. However, the researchers recog­
nised that the technique cannot efficiently detect inter-frame video deepfakes.
Bakas et al. (2021) created a technique [88] that claimed to work well with
compressed videos. However, the performance dropped when one or more
complete group of pictures (GOP) was removed from the video sequence.
Sun et al. (2021) developed a technique [89] that was efficient when applied
to data that had been forged using deepfake technologies. However, while
performing well with data faked with deepfake techniques, the model per­
formed poorly when given neutral data.

3.2.3. Audio deepfakes

Audio manipulation was discussed as one of the latest evolutions of deepfake
exploitation in several papers [12,30,42,90,91], with synthetic voice generation (text-
to-speech) or voice conversion algorithms producing almost undetectable deep­
fake media. Martin (2022) stated, ‘Industry research firm Gartner predicts that within
two years, 20% of all successful account takeover attacks will use deepfakes and
synthetic voice augmentation’ (p 14) [91]. Zhou et al. (2021) suggested that incor­
porating audio and video streams in deepfake detection models would increase the
performance of the model and potentially identify deepfake content in one or both
streams [59]. Masood et al. (2022) noted audio deepfakes had received less atten­
tion than other detection areas, even though synthetic voices were a threat to
100 L. STROEBEL ET AL.

voice-controlled and voice-authenticated systems, presenting a unique challenge

going forward [12]
Khochare et al. (2021) compared two approaches for audio deepfake detec­
tion [42]. A feature-based classification approach utilising machine learning
algorithms, which converted an audio file sample into a collection of features
that were used to calculate whether it was real or fake. Then an image-based
classification approach utilising deep-learning algorithms, where melspectro­
grams produced from a sample and examined by several CNN models deter­
mined if the audio was real or fake. The outcome of this comparison [42] was
that image-based classification (with deep learning algorithms) outperformed
feature-based classification using machine learning. However, they noted [42]
raw audio classifier models would improve accuracy and overhead, as there
would be no pre-processing (manipulation) of the audio files needed.

3.2.4. Datasets
Hazirbas et al. (2021) suggested the top 5 deepfake detection tools per­
form poorly on some specific groups as they did not generalise to all
people [3]. This paper [3] went on to introduce a new dataset, called the
Casual Conversations dataset. It was made up of 45,000 videos using
3,011 subjects from various age, gender and skin tone groups – which
were annotated with additional classifiers to enhance the accuracy of the
AI models trained on it. The conclusion of this study [3] was the deepfake
detection techniques trained on a traditional dataset had a strong biased
towards lighter skin tones, and gender classification was more successful
for older age groups (+45 years old).
Nadimpalli et al. (2022), after examining the gender labelling on current
datasets, concluded when used to train deepfake detection tools they
produced biased results, skewed to successfully identify male-based deep­
fakes more often than female-based ones [85]. That is, female-based
content had a higher false match and higher non-match rate than male-
based content.
Xu et al. (2022) went further, investigating 41 distinct classifiers –
demographic (age, gender, ethnicity) and non-demographic (hair, skin,
accessories, etc), across five deepfake datasets to determine if the data­
sets had diversity. The paper [4] also analysed whether the deepfake
detection tools themselves had AI bias which could lead to generalizabil­
ity, fairness and security issues. The analysis [4] concluded both, the
investigated datasets and deepfake detection models, exhibited bias
with regard to demographic and non-demographic classifiers. That is,
the datasets showed a lack of diversity across the attributes of their
content and the detection models themselves had a strong bias toward
certain attributes.
 JOURNAL OF CYBER SECURITY TECHNOLOGY 101

Almutairi and Elgibreen (2022) identified a lack of datasets available for audio
deepfake detection model training, particularly non-English-based datasets [11].
They also questioned the impact of accents (even in English-based datasets) on
the accuracy of audio deepfake detection [11].
To conclude this section, we would like to mention several papers
[15,92], published in 2022, which presented medical imagery deepfakes
as an emerging threat. Solaiyappan et al. (2022) discussed the need for
more work in detecting deepfakes generated by GANs rather than tamper­
ing methods, such as copy-move or image-splicing methods [15]. However,
Arora et al. (2022) recognised the threat of deepfake images in medical
imagery but also suggested that deepfake generation in this area was an
opportunity to build realistic, effective training material that would not be
impacted by restrictions place on personal information management [92].

4. Observations
In this section, we have addressed the research questions mentioned above,
which we believed were the most relevant and important in this area of research.
The section discusses the deepfake detection techniques identified through this
SLR, including metrics gathered on the effectiveness of these techniques and the
datasets used by them. Advances in deepfake detection technologies were noted
and current challenges and limitations were identified. Finally, potential trends in
future deepfake detection techniques have been discussed.

4.1. Deepfake detection techniques

4.1.1. RQ1 what are the current deepfake detection techniques?

We identified 48 deepfake detection models which met the objectives of this
SLR. These covered deepfake detection on all types of media – image, video,
audio, text and a combination of these. Figure 4 shows the media types
discussed in the documents reviewed. From Table 4, which details the deepfake
detection models and when they were published, we can see there were 28
noteworthy deepfake detection techniques proposed in 2021 and 23 in 2022 –
bearing in mind the search for this SLR was conducted up to and including
August 2022.
Table 6 details the acronyms used in Table 4 – for example: CNN++ repre­
sents convolution neural network extended with other modules/components

4.1.2. RQ1A what is the effectiveness of these techniques?

The metrics shown in Table 4, against the identified deepfake detection techni­
ques, were either Area Under the ROC Curve (AUC) and/or accuracy (ACC). All
models reported high performances in these metrics when trained, validated, and
102 L. STROEBEL ET AL.

tested on the same dataset. However, many experienced a significant perfor­

mance drop when trained on one dataset and validated and tested on another.
Another impact on the performance of these models was the quality of the
data in the dataset used [57] – that is, if the dataset contained high-quality data,
the performance metric was very high, but when the dataset consisted of low-
quality data (for example compressed images), the same model experienced
a drop in performance. Many models [32,41,51] showed comparison metrics
when trained, validated, and tested on several datasets to evaluate performance
across different quality data and this also contributed to the variations in the
metrics shown in Table 4 It was noted [8] that until standardised performance
evaluation criteria for deepfake detection models had been established, the
metrics recorded in the research of individual models had little value, as they
could not be used as a viable comparison metric.
These factors indicated many of the models were not generalisable or robust
when presented with challenging or unknown conditions, meaning they were
not equipped to handle real-world applications [57].

4.1.3. Datasets
There were 30 datasets used across the deepfake detection techniques identi­
fied by this SLR, which have been detailed in Table 5. However, many of the
deepfake detection models used one or more of the following datasets:
FaceForensics++ - contains 1,000 real and 4000 manipulated videos sourced
from 977 YouTube videos. The fake videos come in two compression versions –
high quality (c23) and low quality (c40) [54]. This is a first-generation dataset
that was produced in 2019. It was the most used dataset identified for this SLR,
being used in 51% of the models listed.

Figure 4. Reviewed documents classified by the discussed media.

 JOURNAL OF CYBER SECURITY TECHNOLOGY 103

Table 6. Datasets discussed in the reviewed documents

Ratio
Data Approx Real: %
Name Type Size Fake Gen Year Usage Source
VidTIMIT Audio-Video Dataset VidTIMIT Video/ 350 1:0 1 2001 2 [63,64]
Audio
ImageNet Image 14,000,000 1:0 1 2009 4 [65]
+
Lung Image Database LIDC- Medical 244,527 1:0 1 2012 1 [66]
Consortium image collection IDRI imagery
CelebA Image 202,599 1:0 1 2015 6 [67]
Large-scale Scene LSUN Image 1,000,000 1:0 1 2015 1 [68]
UNderstanding
FaceScrub Image 106,863 1:0 1 2016 1 [69]
Deepfakes Image Dataset DFID Image 19,809 1:0.6 1 2018 1 [47,64]
Unknown-attack Deepfake UADFV Video 98 1:1 1 2018 5 [70]
Video
DeepfakeTIMIT DFTIMIT Video 960 0:1 1 2018 11 [70]
VGGFace2 Image 3,310,000 1:0 1 2018 1 [71]
CASIA-WebFace CASIA Image 494,414 1 2018 4 [72]
FaceForensics++ FF++ Video 5,000 1:4 1 2019 51 [70,73]
100K-Face Image 100,000 0:1 1 2019 1 [60]
ASV spoof 2019 - LA subset ASVS19- Audio 121,461 1:8.7 1 2019 1 [74]
LA
UADFV UADFV Video 98 1:1 1 2018 5 [70]
DeepFakeDetection DFD Video 3,431 1:8.5 2 2019 23 [12,70,73]
Real and Fake Face-Detection RFFD Images 2,041 1:0.8 2 2019 4 [75]
CT-GAN Dataset CT-GAN Medical 100 0:1 2 2019 1 [76]
imagery
Fake or Real FoR Audio 195,000 2 2019 6 [77]
Deepfake Detection Challenge DFDC Video 128,154 1:3.6 2 2020 20 [12,70]
Dataset
CelebDF CDF Video 6,229 1:1.95 2 2020 33 [34,70,78]
DeeperForensics-1.0 DFor-1.0 Video 60,000 5:1 2 2020 2 [73]
Diverse Fake Face Dataset DFFD Image/ 2,600,000 1:1.8 2 2020 1 [79]
Video
Hybrid Fake Face HFF Image 155,000 1:1.5 2 2021 1 [80]
Wild DeepFake WDF Video 7,314 1:0.9 2 2021 1 [34,81]
FaceForensics in the Wild FFIW- Video 10,000 0:1 3 2021 1 [82]
10K
Casual Conversations CC Video 45,186 1:0 3 2021 1 [83]
Korean Deepfake Detection KoDF Images/ 237,942 1:2.6 3 2021 1 [84]
Dataset Video
Gender Balanced DeepFake GBDF Video 10,000 1:4 3 2022 1 [85]
Dataset

CelebDF – contains 590 real videos sourced from YouTube that have been
manipulated into 5,639 deepfake videos. This dataset contains higher-quality
videos than most, as an attempt has been made to remove visible source
artifacts. It was used in 34% of the models listed.
Deepfake Detection – was created by Google and has 3,431 videos, with a ratio
of 1 real to 8.5 fake videos in the dataset. It was used in 23% of the models listed.
Deepfake Detection Challenge Dataset - contains 128,154 videos sourced
from 3,426 paid actors – the video breakdown is 104,500 fake and 23,654 real
videos. This dataset consists of videos in different lighting conditions (indoor/
outdoor) that were taken with high-resolution cameras. It was used in 19% of
the models listed.
104 L. STROEBEL ET AL.

4.2. Deepfake detection advancement

4.2.1. RQ2 what advances have occurred since the previous SLRs?
In an effort to create robust and generalisable deepfake detection models
[31,50], many papers proposed the combination of traditional techniques
(CNN, DNN, LTSM) with other modules to produce ensembled and multi-
attentional architectures. This evolution can be seen via the published dates
of the documents reviewed for this SLR. Stand-alone CNN models and some
multi-attentional architectures were proposed in papers published in 2021, but
only multi-attentional architectures were discussed in the papers published in
2022.
Audio deepfakes were recognised in the literature reviews/surveys reviewed
for this SLR, but mostly they focused on audio-visual deepfakes, rather than
audio alone. Almutairi and Elgibreen (2022) [11] and Masood et al. (2022) [12]
were the first to include audio deepfakes into their literature reviews and both
of these papers were published towards the end of our search date (May and
June 2022 respectively). This indicates audio deepfakes are a new area of
interest. This was also the case with medical imagery deepfakes. The only papers
[15,92] found talking about medical imagery deepfakes were published in June
and July 2022, which again suggested it was an emerging area of research.
Traditionally, deepfake detection models focused on only one type of facial
manipulation – that is, face swapping or feature manipulation of the primary
face in the image or video. However, recent models [39,61] were used to assess
the whole image for manipulation, including images with no faces in them,
which suggested another emerging area of deepfake detection. One model [44]
went so far as successfully identifying multiple deepfake faces in images over
frame by frame examination of videos.

4.3. Challenges and limitations

4.3.1. RQ3 what are the current challenges involved in deepfake detection?
Some of the main challenges identified with deepfake detection technologies
were:

● The quality, fairness, and trust of deepfake datasets (biased and imbal­
anced data) [6,7,9,12,90];
● Robustness of the deepfake detection techniques against unknown attack
types [7];
● Temporal aggregation [7,12]; and
● Social media laundering [12]

This survey [8] noted a need for competitive baselines to evaluate deepfake
detection techniques so the true performance could be assessed. Zhang (2022)
 JOURNAL OF CYBER SECURITY TECHNOLOGY 105

agreed, indicating the deepfake detection techniques up to 2020, reviewed in

this survey [6], were not robust and there were many challenges when assessing
generation and detection models, including a lack of benchmark test methods
and quality datasets.
In addition, deepfake detection techniques, due to a lack of generalisation or
their narrow field of specialisation, offered poor practical applications. An
example of this was discussed by Guo et al. (2021) where an adaptive manipu­
lated traces extraction network (AMTEN) was combined with a CNN to enhance
the efficiency of the model, but was found to suffer from a lack of generalisa­
bility to make it useful in most practical settings [31]. Similarly, the 3D CNN
model proposed by Liu et al. (2021) failed to be of any practical use due to the
consumption of resources it required [24].
Concern over the viability of datasets was discussed in many papers
[3,4,6–9,12,85,90] with several areas identified as needing further research
and improvement.
Xu et al. (2022) proposed that most research in this area had focused on
developing deepfake detection solutions utilising the available datasets, but
generalizability issues had arisen from training these solutions on less diverse
datasets [4]. Furthermore, they found that the detection solutions developed
inappropriate bias due to the presence of certain attributes in the datasets (for
example, being male or black or having a big nose). This was also noted by
Masood et al. [12] who suggested there was an urgent need to address data
fairness in detection techniques.
Hazirbas et al. (2021) produced a new, ‘balanced’ dataset using subjects from
various age, gender and skin tone groups and annotated with additional
classifiers to enhance the accuracy of deepfake detection techniques trained
on it [3]. The age and gender classifiers in the dataset were annotated by the
subjects themselves, which was believed would produce an unbiased view of
these classifiers. The skin tone classifier was annotated using the Fitzpatrick skin
phototype scale. This study [3] found the deepfake detection techniques exam­
ined had a strong biased towards lighter skin tones. This was concluded
because they mostly failed on dark skin tones. Also, gender classifications
were more successful for older age groups (+45 years old) across all skin tones.
As mentioned in section 4.1, the SLR was unable to identify any models that
conclusively showed a sustained, high level of capability across multiple data
qualities, that is, high and low quality data, particularly when presented with
challenging or unknown conditions like light variations or biometric features
(such as glasses or heavy eyebrows). Nor could any models detect deepfakes
from multiple forms of manipulation (that is, face swapping and facial manip­
ulation) especially when the manipulation was unknown to the model. This
meant current deepfake detection models were not generalisable or robust
when presented with challenging or unknown conditions, and were not
equipped to handle real-world applications [57].
106 L. STROEBEL ET AL.

Temporal aggregation, the temporal consistency between frames in deep­

fake video content – that is, real and fake content appearing in sequential
intervals that would require extra effort to identify and classify, was discussed
in [7,12] as an area as yet to be investigated.
Images appearing on social media platforms have been heavily compressed
during the upload process in order to limit the size and bandwidth needed.
These manipulations, called social media laundering [12], remove a lot of the
artifacts that have been added to an image or video during the deepfake
generation process. This means the evidence that would normally aid deepfake
detection is no longer available and makes detecting deepfake content from
social media platforms challenging.
We had one final comment on deepfake detection challenges related to the
potential for deepfake usage in legal cases. Media evidence is already promi­
nent in the courtroom; and it would be surprising if a courtroom trial did not
have any picture, video or audio evidence admitted either on behalf of the
defence or prosecution. In 2016, the Bureau of Justice Assistance, U.S.
Department of Justice estimated that 80% of crimes had video evidence asso­
ciated with them. It seems logical then that deepfake videos, audio and imagery,
would begin to be used by either side in a court of law, whether it be intentional
or accidental. Pfefferkorn (2016) talks about a concept known as the ‘reverse CSI
effect’ in which those involved in the legal process, jurors for example, begin to
doubt it is ever possible to determine what is real [93]. Pfefferkorn believes that
the courts already have the tools required to be able to weed out inauthentic
evidence, although it may involve extra effort [93]. Venema (2020) believes that
the courts are at risk of falling victim to the ‘deepfake defence’ if they haven’t
already [94]. The future challenge lies in keeping up with the pace of deepfakes
and being able to accurately detect them.

4.4. Looking ahead

4.4.1. RQ4 can any future trends be identified, and if so, what are they?
The majority of documents reviewed for this SLR related to image, video and
image/video deepfake detection (90%) with the remaining discussing techni­
ques for audio, audio/video or a combination of all media. This showed little
research had been directed at deepfakes other than image and video – which
suggests the main driver for deepfake detection has been social media con­
glomerates, which have a massive footprint in these media.
AI-generated voices have now become so ‘life-like’ it is difficult to distinguish
them from real audio. The original need for audio deepfakes used in digital
assistant devices has resulted in tools becoming publicly available that can
produce undetectable audio deepfakes from a very small source sample.
These are being touted as a growing threat to organisations. That is, real-time
audio deepfake has the potential to be the new social engineering tool for
 JOURNAL OF CYBER SECURITY TECHNOLOGY 107

cyber-attacks. As email phishing attacks are mitigated with awareness training,

spear phishing attacks using audio impersonation are becoming the new age of
deepfake threats.
Identity compromise via voice activation systems is also a serious and grow­
ing cyber security threat. The use of voice biometrics is expected to increase in
the next few years, as organisations move to passwordless authentication, and
audio deepfakes present a challenge that needs to be overcome sooner rather
than later.
There is a general lack of datasets for audio deepfake detection [11,12],
particularly non-English speaking datasets, and this means a lot of work is
needed in this area. Also, as yet, no research has been undertaken to evaluate
the impact of accents on successful deepfake detection [11]. As already men­
tioned, audio deepfakes are a growing threat and immediate attention in these
areas is required if deepfake detection models are to effectively manage the
threat landscape.
More work also needs to be done on annotating datasets and ensuring
detection models remain unbiased through and after training. This will go
some way to reducing the bias found in deepfake detection techniques but
work still needs to be done to ensure the techniques themselves are free from
bias.

5. Conclusion
This paper comprehensively discussed methods used in the detection of deep­
fakes and determined current techniques. As part of the review process, this
paper categorised papers into two categories, a) Previous Literature Reviews
and Surveys and b) modern deepfake detection methods. The SLR reviewed 8
literature review/survey documents with 7 of those published in 2022. It also
reviewed 75 documents from category b) modern deepfake detection methods,
ranging from 2021 to August 2022, which identified developing trends and
future research areas. The majority (90%) of the papers reviewed focused on
video and image or a combination of both.
This paper identified 48 noteworthy detection models, across all types of
media including image, video, audio, text, and a combination of all. Of these 48,
23 detection techniques were from research and proposals from 2022 leading
up to August 2022, with a further 28 proposed in 2021.
Advancements in deepfake detection have proposed a combination of tradi­
tional techniques (CNN, DNN and LTSM) and other modules to strengthen these
techniques, producing ensembled and multi-attentional architectures. Recent
papers written have adopted this strategy with many suggesting multi-
attentional architectures in 2021 papers and papers written in 2022 only sug­
gesting multi-attentional architectures.
108 L. STROEBEL ET AL.

Models discussed within this paper were judged against Area Under the ROC
Curve (AUC) and/or accuracy (ACC); a growing method of reporting the effec­
tiveness of deepfake techniques. It is worth noting that all models this report
discovered, reported high performance in these metrics when trained, vali­
dated, and tested on the same dataset. However, when these models were
trained and validated on differing datasets there were noticeable performance
drops. As such this report concluded models were not robust or adaptable when
presented with challenging or unknown conditions.
As part of this review, information was discovered that suggested deepfake
detection tools perform poorly on various cultural and ethnic groups. It was
concluded in studies reviewed that deepfake detection techniques trained on
a standard dataset had a strong bias towards lighter skin tone and gender
classification and had greater success with people aged above 45 years. As
such, challenges identified within this paper include the quality, fairness, and
trust of the deepfake datasets. Also, the robustness of these datasets when
being used against unknown attack types was a challenge for detection. Many
techniques reviewed have a large reliance on specific datasets, if these datasets
do not represent a broad section of society, including cultural or ethnic repre­
sentation, their effectiveness is severely reduced. As such, the viability of data­
sets is a large challenge moving forward.
A recommendation this paper suggests is the need to create a standardised
approach to dataset usage and a uniform rating system where techniques can
be validated against each other, reducing the ability to select the best or most
desirable datasets. Currently, deepfake detection technique creators and own­
ers can elevate their models using targeted datasets known to produce better
results for that model type, making it appear superior.
There are current shortfalls within the research field of deepfake detection
outside of the current popular image and video types. As mentioned, this paper
found the majority of research reviewed related to one or both types. As a result,
more research is required in additional fields including audio-only deepfake
detection.
Future trends identified in this paper include audio deepfakes as the latest
evolution of deepfake exploitation, utilising synthetic voice generation or voice
conversation algorithms that output undetectable deepfake audio. As pre­
viously mentioned, Gartner has predicted, within two years 20% of all successful
account takeovers will be deepfakes and synthetic voice augmentation [91].
Leading to audio deepfakes being used in deception methods including spear
phishing scams. This prediction highlights a future trend in deepfake generation
and the requirement for additional research in audio deepfake detection. In
addition to audio deepfakes, several papers also mentioned concerns surround­
ing medical imagery deepfakes. As such, additional research is required in
detection methods of these deepfakes types.
 JOURNAL OF CYBER SECURITY TECHNOLOGY 109

This research has highlighted a positive aspect of deep fake material. Such as
deep fake imagery can be used for good during medical training providing
realistic material.
With the evolution of technology and the low-cost barriers to entry, the
advancement of deep fakes will progress with a rapid trajectory. As this evolves,
future challenges in detection techniques will be required to adapt at the same
level or greater to those technological advancements.

Disclosure statement
No potential conflict of interest was reported by the authors.

ORCID
Mohiuddin Ahmed http://orcid.org/0000-0002-4559-4768

References
[1] Curry L, Nembhard I, Bradley E. Qualitative and mixed methods provide unique con­
tributions to outcomes research. Circulation. 2009;119(10):1442–1452.
[2] Abdulreda AS, Obaid AJ. A landscape view of deepfake techniques and detection
methods. Int J Nonlinear Anal Appl. 2022;13(1):745–755.
[3] Hazirbas C, Bitton J, Dolhansky B. Towards Measuring Fairness in AI: the Casual
Conversations Datase. IEEE Transactions on Biometrics, Behavior, and Identity
Science. 2021.
[4] Xu Y, Terhörst P, Raja K, et al. A Comprehensive Analysis of AI Biases in DeepFake
Detection with Massively Annotated Databases. arXiv preprint arXiv:2208.05845, 2022.
[5] Weerawardana M, Fernando T, Deepfakes detection methods: a literature survey, in 2021
10th International Conference on Information and Automation for Sustainability (ICIAfS).
2021, IEEE Access: Negambo, Sri Lanka. p. 76–81.
[6] Zhang T. Deepfake generation and detection, a survey. Multimedia Tools Appl. 2022;81
(5):6259–6276.
[7] Malik A, Kuribayashi M, Abdullahi SM, et al. DeepFake detection for human face images
and videos: a survey. IEEE Access. 2022;10:18757–18775.
[8] Juefei-Xu F, Wang R, Huang Y, et al. Countering malicious deepfakes: survey, battle­
ground, and horizon. Int J Comput Vis. 2022;130(7):1678–1734. DOI:10.1007/s11263-
022-01606-8
[9] Rana MS, Nobi MN, Murali B, et al. Deepfake detection: a systematic literature review.
IEEE Access. 2022;10:25494–25513.
[10] Celebi N, Liu Q, Karatoprak M, A survey of deep fake detection for trial courts, in 9th
International Conference on Artificial Intelligence and Applications (AIAPP 2022). 2022,
ResearchGate: Vancouver, Canada. p. 227–238.
[11] Almutairi Z, Elgibreen H. A review of modern audio deepfake detection methods:
challenges and future directions. Algorithms. 2022;15(5):155.
[12] Masood M, Nawaz M, Malik KM, et al. Deepfakes generation and detection: state-of-the-
art, open challenges, countermeasures, and way forward. Appl Intell. 2022;53
(4):3974–4026. DOI:10.1007/s10489-022-03766-z
110 L. STROEBEL ET AL.

[13] Yang J, Xiao S, Li A, et al. MSTA-Net: forgery detection by generating manipulation

trace based on multi-scale self-texture attention. IEEE Trans Circuits Syst Video Technol.
2022;32(7):4854–4866. DOI:10.1109/TCSVT.2021.3133859
[14] Mikhaltsov VE, Semenova ZV, Stepanova EA, Effectiveness analysing of modification of
training data set for voice forgery detection system. Journal of Physics: Conference
Series. 2022;2182(1):012096.
[15] Solaiyappan S, Wen Y. Machine learning based medical image deepfake detection:
a comparative study. Mach Learn Appl. 2022;8:100298.
[16] Fang S, Wang S, Ye R. DeepFake video detection through facial sparse optical flow
based light cnN. Journal of Physics: Conference Series. 2022;2224(1):012014.
[17] Fu T, Xia M, Yang G. Detecting GAN-generated face images via hybrid texture and
sensor noise based features. Multimedia Tools Appl. 2022;81(18):26345–26359.
[18] Lu Y, Liu Y, Fei J, et al. Channel-wise spatiotemporal aggregation technology for face
video forensics. Secur Commun Networks. 2021;2021:1–13.
[19] Raskar PS, Shah SK. Real time object-based video forgery detection using YOLO (V2).
Forensic Science International. 2021;327.
[20] Ramachandran S, Nadimpalli AV, Rattani A, An experimental evaluation on deepfake
detection using deep face recognition, in 2021 International Carnahan Conference on
Security Technology (ICCST). 2021, IEEE: Hatfield, United Kingdom. p. 1–6.
[21] Shad HS, Rizvee MM, Roza NT, et al. Comparative analysis of deepfake image detection
method using convolutional neural network. Comput Intell Neurosci.
2021;2021:3111676.
[22] Luo Y, Ye F, Weng B, et al. A novel defensive strategy for facial manipulation detection
combining bilateral filtering and joint adversarial training. Secur Commun Networks.
2021;2021:1–10.
[23] Hsu H-W, Ding J-J, Deepfake algorithm using multiple noise modalities with two-branch
prediction network, in 2021 Asia-Pacific Signal and Information Processing Association
Annual Summit and Conference (APSIPA ASC). 2021, APSIPA: Tokyo, Japan. p. 1662–1669.
[24] Liu J, Zhu K, Lu W, et al. A lightweight 3D convolutional neural network for deepfake
detection. Int J Intell Syst. 2021;36(9):4990–5004. DOI:10.1002/int.22499
[25] Nguyen XH, Tran TS, Le VT, et al. Learning spatio-temporal features to detect manipulated
facial videos created by the deepfake techniques. Forensic Sci Int: Digital Invest.
2021;36:301108.
[26] Lee S, Tariq S, Shin Y, et al. Detecting handcrafted facial image manipulations and
GAN-generated facial images using shallow-FakeFaceNet. Appl Soft Comput.
2021;105:107256.
[27] Kohli A, Gupta A. Detecting DeepFake, FaceSwap and face2face facial forgeries using
frequency CNN. multimedia tools and applications. Int J. 2021;80(12):18461–18478.
[28] Luo Z, Kamata S-I, Sun Z, Transformer and node-compressed dnn based dual-path system
for manipulated face detection, in 2021 IEEE International Conference on Image
Processing (ICIP). 2021, Anchorage, AK, USA. p. 3882–3886.
[29] Ru Y, Zhou W, Liu Y, Sun J, Li Q. Bita-net: bi-temporal attention network for facial video
forgery detection. 2021 IEEE International Joint Conference on Biometrics (IJCB),
Shenzhen, China. 2021;1–8. doi:10.1109/IJCB52358.2021.9484408
[30] Zhang M. FST-Net: exploiting Frequency Spatial Temporal Information for Low-Quality
Fake Video Detection. in 2021 IEEE 33rd International Conference on Tools with Artificial
Intelligence (ICTAI). 2021. Washington, DC, USA: IEEE.
[31] Guo Z. Fake face detection via adaptive manipulation traces extraction network.
Computer Vision and Image Understanding. 2021;204.
 JOURNAL OF CYBER SECURITY TECHNOLOGY 111

[32] Chen B, Tan S, Alazab M. FeatureTransfer: unsupervised domain adaptation for

cross-domain deepfake detection. Secur Commun Networks. 2021;2021:1–8.
[33] Ananthi M. A secure model on Advanced Fake Image-Feature Network (AFIFN) based on
deep learning for image forgery detection. Pattern Recognit Lett. 2021;152:260–266.
[34] Zhou Y. Face Forgery Detection Based on Segmentation Network, in 2021 IEEE International
Conference on Image Processing (ICIP). 2021, IEEE: Anchorage, AK, USA. p. 3597–3601.
[35] Sun Z. Improving the efficiency and robustness of deepfakes detection through precise
geometric features, in Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition. 2021, Nashville, TN, USA. p. 3609–3618.
[36] Shichao G, Ming X, Gaobo Y. Gaobodual-tree complex wavelet transform-based direction
correlation for face forgery detection. Secur Commun Networks. 2021;2021:1–10.
[37] Gong D, Kumar YJ, Goh OS, et al. DeepfakeNet, an efficient deepfake detection method.
Int J Adv Comput Sci Appl. 2021;12(6).
[38] Jin X, Ye D, Chen C. Countering spoof: towards detecting deepfake with multidimen­
sional biological signals. Secur Commun Networks. 2021;2021:1–8.
[39] Singh B, Sharma DK. SiteForge: detecting and localizing forged images on microblog­
ging platforms using deep convolutional neural network. Comput Ind Eng.
2021;162:107733.
[40] Yang J, Xiao S, Li A, et al. Detecting fake images by identifying potential texture
difference. Future Gener Comput Syst. 2021;125:127–135.
[41] Xu Z, Liu J, Lu W, et al. Detecting facial manipulated videos based on set convolutional
neural networks. J Vis Commun Image Represent. 2021;77:103119.
[42] Khochare J, Joshi C, Yenarkar B, et al. A deep learning framework for audio deepfake
detection. Arab J Sci Eng. 2021;47(3):3447–3458. DOI:10.1007/s13369-021-06297-w
[43] Wang Z, Yiwen G, Zuo W. Deepfake forensics via an adversarial game. IEEE Trans Image
Process. 2022;31:3541–3552.
[44] Deng L, Suo H, Li D. Deepfake video detection based on efficientnet-v2 network.
Comput Intell Neurosci. 2022;2022:3441549.
[45] Wang G, Jiang Q, Jin X, et al. MC-LCR: multimodal contrastive classification by locally
correlated representations for effective face forgery detection. Knowledge-Based Syst.
2022;250:109114.
[46] Wang T, Liu M, Cao W, et al. Deepfake noise investigation and detection. forensic science
international: digital investigation. Forensic Sci Int: Digital Invest. 2022;42:301395.
[47] Ganguly S, Ganguly A, Mohiuddin S, et al. ViXNet: vision transformer with xception
network for deepfakes based video and image forgery detection. Expert Syst Appl.
2022;210:118423.
[48] Cozzolino D. ID-Reveal: identity-aware DeepFake Video Detection, in 2021 IEEE/CVF
International Conference on Computer Vision (ICCV). 2021, IEEE: Montreal, QC, Canada. p.
15088–15097.
[49] Pu W, Hu J, Wang X, et al. Learning a deep dual-level network for robust DeepFake
detection. Pattern Recogn. 2022;130:108832.
[50] Taeb M, Chi H. Comparison of deepfake detection techniques through deep learning.
J Cybersecur Privacy. 2022;2(1):89–106.
[51] Wang J, Sun Y, Tang J. LiSiam: localization invariance siamese network for deepfake
detection. IEEE Trans Inf Forensics Secur. 2022;17:2425–2436.
[52] Kolagati S, Priyadharshini T, Mary Anita Rajam V. Exposing deepfakes using a deep
multilayer perceptron – convolutional neural network model. Int J Inf Manage Data
Insights. 2022;2(1):100054.
112 L. STROEBEL ET AL.

[53] Hu J, Liao X, Wang W, et al. Detecting compressed deepfake videos in social networks
using frame-temporality two-stream convolutional network. IEEE Trans Circuits Syst
Video Technol. 2022;32(3):1089–1102. DOI:10.1109/TCSVT.2021.3074259
[54] Yu M, Ju S, Zhang J, et al. Patch-DFD: patch-based end-to-end DeepFake discriminator.
Neurocomputing. 2022;501:583–595.
[55] Chen B, Li T, Ding W. Detecting deepfake videos based on spatiotemporal attention
and convolutional LSTM. Inf Sci. 2022;601:58–70.
[56] Saif S. Generalized Deepfake Video Detection Through Time-Distribution and Metric
Learning. IT Prof. 2022;24(2):38–44.
[57] Chamot F, Geradts Z, Haasdijk E. Deepfake forensics: cross-manipulation robustness of
feedforward- and recurrent convolutional forgery detection methods. Forensic Sci Int:
Digital Invest. 2022;40:301374.
[58] Vamsi VVVNS. Deepfake detection in digital media forensics. Global Transitions
Proceedings. 2022;3(1):74–79.
[59] Zhou Y, Lim S-N, Q.C.C.O.O. Ieee/Cvf International Conference on Computer Vision
Montreal, Joint Audio-Visual Deepfake Detection, in 2021 IEEE/CVF International
Conference on Computer Vision (ICCV). 2021, Montreal, Canada: IEEE. p. 14780–14789.
[60] Prototypr. AI-Generated Faces: Free Resource of 100K Faces Without Copyright. 2019 18
September, 2019 [cited 08/10/2022]; Available from: https://blog.prototypr.io/gener
ated-photos-free-resource-of-100k-diverse-faces-generated-by-ai-2144a8615d1f.
[61] Liu X. PSCC-Net: progressive Spatio-Channel Correlation Network for Image
Manipulation Detection and Localization. IEEE Transactions on Circuits and Systems
for Video Technology. 2022;1.
[62] Coccomini DA, Messina N, Gennaro C, et al. Combining EfficientNet and vision
transformers for video deepfake detection. In: Sclaroff S, Distante C, Leo M,
Farinella GM Tombari F, editors. Image analysis and processing – iciap 2022. iciap
2022. lecture notes in computer science. Vol. 13233. Cham: Springer; 2022. doi:10.
1007/978-3-031-06433-3_19.
[63] Babbar A The VidTIMIT Audio-Video Dataset. 2016 [cited 2022 18 September]; Available
from: https://www.kaggle.com/datasets/akshay4/speakerrecognition .
[64] Afchar D. MesoNet: a Compact Facial Video Forgery Detection Network, in 2018 IEEE
International Workshop on Information Forensics and Security (WIFS). 2018, Hong Kong:
IEEE. p. 1–7.
[65] University S ImageNet. 2021 [cited 2022 18 September]; Available from: https://www.
image-net.org/.
[66] Bilello E LIDC-IDRI. 2022 16 September, 2022 [cited 2022 20 September]; Available from:
https://wiki.cancerimagingarchive.net/display/Public/LIDC-IDRI.
[67] Li J CelebFaces Attributes (CelebA) Dataset. 2018 [cited 2022 18 September]; Available
from: https://www.kaggle.com/datasets/jessicali9530/celeba-dataset .
[68] Yu F. LSUN: construction of a large-scale image dataset using deep learning with
humans in the loop. Vol. 9. 2015. 10.48550/arXiv.1506.03365.
[69] Group VI FaceScrub. 2021 23 March 2021 [cited 2022 19 September]; Available from:
http://vintage.winklerbros.net/facescrub.html.
[70] Dolhansky B. The DeepFake Detection Challenge (DFDC) Dataset. arXiv preprint
arXiv:2006.07397, 2020.
[71] Xie W VGGFace2 dataset for face recognition. 2020 18 February, 2020 [cited 2022 19
September]; Available from: https://github.com/ox-vgg/vgg_face2.
[72] Code PW CASIA-Webface. [cited 2022 19 September]; Available from: https://papers
withcode.com/dataset/casia-webface.
 JOURNAL OF CYBER SECURITY TECHNOLOGY 113

[73] Rathgeb C. Handbook of Digital Face Manipulation and Detection. In: Kang SB, editor.
Advances in Computer Vision and Pattern Recognition. 1. Springer International
Publishing. IX; 2022. p. 487.
[74] Zhang Y, Jiang F, Duan Z. One-class learning towards synthetic voice spoofing
detection. IEEE Signal Process Lett. 2021;28:937–941.
[75] University Y Real and Fake Face Detection. 2019 [cited 2022 19 September]; Available
from: https://www.kaggle.com/datasets/ciplab/real-and-fake-face-detection.
[76] Mirsky Y. {CT-GAN}: malicious Tampering of 3D Medical Imagery using Deep Learning. in
28th USENIX Security Symposium (USENIX Security 19). 2019, Santa Clara, CA.
[77] Reimao R. FoR: a Dataset for Synthetic Speech Detection, in 2019 International Conference
on Speech Technology and Human-Computer Dialogue (SpeD). 2019, Timisoara,
Romania: IEEE. p. 1–10.
[78] GitHub I Celeb-DF: A Large-scale Challenging Dataset for DeepFake Forensics. 2021 3 June
2021 [cited 2022 08 October]; Available from: https://github.com/yuezunli/celeb-
deepfakeforensics.
[79] Hao D. DFFD: Diverse Fake Face Dataset. 2020 June 2020 [cited 2022 8 October];
Available from: http://cvlab.cse.msu.edu/dffd-dataset.html.
[80] GitHub I Hybrid Fake Face Dataset. 2021 17 March, 2021 [cited 2022 08 October];
Available from: https://github.com/EricGzq/Hybrid-Fake-Face-Dataset.
[81] Zi B. WildDeepfake: a challenging real-world dataset for deepfake detection. in Proceedings
of the 28th ACM international conference on multimedia. 2020, Seattle, WA, USA.
[82] Zhou T Face Forensics in the Wild. 2021 4 October, 2021 [cited 2022 19 September];
Available from: https://github.com/tfzhou/FFIW.
[83] MetaAI. Casual Conversations Dataset. 2021 [cited 2022 20 September]; Available from:
https://ai.facebook.com/datasets/casual-conversations-dataset/.
[84] Kwon P. KoDF: a Large-scale Korean DeepFake Detection Dataset, in International
Conference on Computer Vision (ICCV). 2021, IEEE: Montreal, Canada. p. 10724–10733.
[85] Nadimpalli AV, Rattani A, GBDF: gender balanced deepfake dataset towards fair deepfake
detection, in International Conference on Pattern Recognition 2022. 2022, ResearchGate:
Montreal, Canada.
[86] Zhang J, Ni J, Xie X, DeepFake videos detection using self-supervised decoupling network,
in 2021 IEEE International Conference on Multimedia and Expo (ICME). 2021 : Shenzhen,
China. p. 1–6.
[87] Chen B, Liu X, Zheng Y, et al. A robust gan-generated face detection method based on
dual-color spaces and an improved xception. IEEE Trans Circuits Syst Video Technol.
2022;32(6):3527–3538. DOI:10.1109/TCSVT.2021.3116679
[88] Jamimamul Bakas RN, Bakshi S. Detection and localization of inter-frame forgeries in
videos based on macroblock variation and motion vector analysis. Comp Elec Eng.
2021;89:106929.
[89] Fang Sun NZ, Pan X, Song Z. Deepfake detection method based on cross-domain
fusion. Secur Commun Networks. 2021;2021:11.
[90] Muneef Z, Elgibreen H. A review of modern audio deepfake detection methods:
challenges and future directions. Algorithms. 2022;15(5):19.
[91] Martin EJ. Deepfakes: the latest trick of the tongue. Speech Technology. 2022;27(2):12–16.
[92] Arora A, Arora A. Generative adversarial networks and synthetic patient data: current
challenges and future perspectives. Future Healthc J. 2022;9(2):190–193.
[93] Pfefferkorn R. “Deepfakes” in the courtroom. Public Interest Law J. 2020;29:245–275.
[94] Venema AE, Geradts ZJP. Digital forensics, deepfakes, and the legal process. Sci Tech
Lawyer. 2020;16(4):14–17,23.