Journal of Information Systems and Telecommunication http://jist.acecr.

org
Vol.12, No.1, January-March 2024, 41-48 ISSN 2322-1437 / EISSN:2345-2773

1.
Department of Electrical and Computer Engineering, Tarbiat Modares University of Tehran, Tehran, Iran,
2.
DirectorateYouth and Sport of Babylon, Iraq.

Received: 1 Feb 2023/ Revised: 10 May 2023/ Accepted: 07 Jul 2023

Abstract
Today, social networks have become a prominent source of news, significantly altering the way people obtain news from
traditional media sources to social media. Alternatively, social media platforms have been plagued by unauthenticated and
fake news in recent years. However, the rise of fake news on these platforms has become a challenging issue. Fake news
dissemination, especially through visual content, poses a significant threat as people tend to share information in image
format. Consequently, detecting and combating fake news has become crucial in the realm of social media. In this paper,
we propose an approach to address the detection of fake image news. Our method incorporates the error level analysis
(ELA) technique and the explicit convolutional neural network of the EfficientNet model. By converting the original image
into an ELA image, it is possible to effectively highlight any manipulations or discrepancies within the image. The ELA
image is further processed by the EfficientNet model, which captures distinctive features used to detect fake image news.
Visual features extracted from the model are passed through a dense layer and a sigmoid function to predict the image type.
To evaluate the efficacy of the proposed method, we conducted experiments using the CASIA 2.0 dataset, a widely adopted
benchmark dataset for fake image detection. The experimental results demonstrate an accuracy rate of 96.11% for the
CASIA dataset. The results outperform in terms of accuracy and computational efficiency, with a 6% increase in accuracy
and a 5.2% improvement in the F-score compared with other similar methods.

Keywords: Fake News; EfficientNet; Fake Image; Social Media; Error Level Analysis.

propaganda that consists of deliberate misinformation or

hoaxes spread through traditional print and broadcast
1- Introduction media or online social media fake news[1],[2]. Social
media is among the widely accepted platforms globally. Its
In the recent decade, It has greatly facilitated the rise of characteristics are ease-of-use, rapid rate, and low cost,
social media, the common way for people to get in touch making it the most welcoming online platform for
with each other and share information. Social media has information sharing and social interaction [2],[3]. Today,
many positive characteristics, like ease of use, low cost, more than two-thirds of American adults can access online
and 7x24 information accesses. Unfortunately, fake news news outlets. This increasing rate has made the Internet an
has greatly increased on social platforms. The increasing ideal channel for fake news dissemination. Social media is
rate of fake news has become one of the troubling issues the primary media for the propagation of fake news and,
since it can mislead people. Online fake news tends to be consequently, has been one of the prominent studied areas
diverse and intrusive regarding topics, platforms, and by industries and universities.
styles. Fake news could have many negative impacts on There are different social media platforms with distinctive
individuals, business, and society. So, it is crucial to features. The most popular ones are Facebook and Twitter,
introduce and launch a system that could detect, explore that have been found as the primary sources of fake news
and interpret fake news on social media. It is challenging diffusion [4]. The significant difference between the two is
to come up with a definition for “fake news” that could be that each post on Twitter, a microblogging site, is limited
accepted in general. According to Stanford University, to 380 characters, while on Facebook, the limit is nearly
fake news is “news articles that are intentionally and 60,000 [5],[6].
verifiably false and could mislead readers.” According to The study in [7] indicates that 54% of users in industrial
online Wikipedia, “a type of yellow journalism or countries are worried about “what is real or fake” on social

 Nasrollah Moghaddam Charkari

[email protected]
42
Khalid Zamil1& Moghaddam Charkari, Image Fake News Detection Using Efficient NetB0 Model

platforms. This concern has become more significant after The rest of this paper is structured as follows: Section 2
the 2016 U.S. presidential election [8],[9] due to the presents some of the interesting related work on fake news
influence of social media on political polarization and detection. In section 3, we discuss the methodology.
conflicts among the political parties during the campaign Section 4 present the dataset and experimental results.
period [10]. Another instance is the surge in online fake Finally, the conclusion and feature work are discussed in
news following the lookdown measures to curtail the section 5.
spread of COVID-19 disease. A study recently reported a
25% increase in social media users following the global
lockdown. According to a UNESCO report, “during this 2- Related Works
coronavirus pandemic. Hence, the WHO described all
misinformation related to COVID-19 is often referred to as Social media has evolved into a crucial source of
an "infodemic," which they defined as an overabundance information and an integral aspect of our daily life. The
of offline and online information[10],[3], [12]. To curtail majority of information on social media is in the form of
this menace, many fact-checking online systems such as photographs. Meanwhile, phony news events have been
FactCheck.org have recently come up to verify political increasingly distributed on social media, leading to user
news; however, the practicability of these systems is confusion. The existing news verification techniques rely
restricted due to the numerous” types and formats of fake on features collected from the text content of tweets,
news that facilitate its dissemination on the social network whereas image features are frequently overlooked for
[7], [13] verification of news. Fake news detection on photos has
Fake images in the news play an outstanding part. Fake been the subject of few studies. The absence of training
images are often used to provoke public anger and gather data is one of the drawbacks of using visual-based
public opinion. When it shares in serious repercussions features. Building a human-labeled a fake news dataset is
such as mass killings and religious conflicts, it has an even time-consuming and labor-intensive. As a result, creating a
more devastating impact. Various software tools usually fake news dataset with images or videos to train is
modify fake images. Since, they might severely affect considerably a complex task. The following are the most
people's thoughts. Adobe's state of the content survey recent studies on images in the field of fake news
revealed that engagement for posts with images is three identification [15], [16], [17]
times more destructive than posts with text only. As a Dinesh Kumar Vishwakarma et al. [15]proposed an image-
result, the fake images inside fake news has been increased based fake news detection method. The method comprises
in social media in recent years. So, developing solutions to four core components: “image text extraction, entity
discover fake images and text content on social media extractor, web processing, and processing unit.” Initially,
platforms is a crucial task [3]. Moreover, online social data an algorithm was employed to extract the text region, and
is time-sensible, meaning it appears in a real-time type and then the text was recovered from photos using optical
represents current events and issues. There is an urgent character recognition (OCR). This way, results are fetched
necessity for early detection approaches of fake news from and further classified as reliable or unreliable connections.
the huge number of news articles published daily[14]. The high classification rate for this method is 85%. The
In this article, we propose an approach for fake image dataset included the Google image/ Kaggle / Onion
detection. The main advantages of the proposed method dataset. Zhiwei Jin et al. [18]proposed a method to detect
are as follow; fake images based on visual and statistical image features;
 Computational time: Using efficientNetB0 model the gain ratio method was used to remove redundant
helps us to learn image features with fewer features. This procedure selects 11 elements from 42
parameters compared to other deep learning features. Four classification models, SVM, LR, KStar, and
approaches. Consequently, it leads to find fake RF, have been employed to train the method. The dataset
images in lower computational time, which is a comprises 50,287 tweets and 25,953 images of fake and
crucial task in this area. actual news events on SinaWeibo. The highest accuracy
 Proper feature extraction: Additionally, converting rate was 83.6% using the Random Forest classifier.
the original image into an error-level analysis Francesco et al. [19] proposed a fake image detection
(ELA) image enables the model to capture the method that relies on GAN-based image to image
manipulated features that further daal to translation; this method relies on the modern approaches
effectively detect fake images. taken from the image forensic. CNN has been used to train
 Improved Efficiency: The results of the data. An accuracy rate of 89.03% was reported using
experiments on popular datasets indicate that the 36302 image dataset.
proposed approach outperforms the current state- D. Mangal [16] presented a Multi-Domain Visual Neural
of-the-art methods regarding precision, recall, and Network (MVNN) model for the detection of fake news;
accuracy rates. the model is comprised of “a frequency domain sub-
 43
Journal of Information Systems and Telecommunication, Vol.12, No.1, January-March 2024

network, a pixel domain sub-network, and a fusion sub- 3- Proposed Method

network.” The fusion sub-network fuses the obtained
feature vectors from the pixel and frequency domain sub- We propose a method to deal with the problem of image
network through a fully connected layer; SoftMax fake news identification in this paper.
activation is utilized to project the vector into either fake- The overall framework of the proposed method is
news images or actual news. The Weibo dataset has been displayed in fig1 and algorithm1.
used in the experiments, and the accuracy rate reached to
84.6%.
Singh et al. [20] proposed an image-based fake news
detection method. CNN with an attention mechanism is
employed to detect fake images over the social network.
The model uses high-pass filters to the kernel weights of
the NN initialization. The two-dataset from Twitter and
The approach described in [18], used a statistical and
machine learning methods, suffers from efficiently fake
news detection due to the challenge of identifying
manipulated features in handcrafted image features. This
limitation results in poor model generalization. As
described in [22], the forensic methods have been Fig. 1. The overall framework of the proposed method.
employed to extract image features. However, these
features are specific to particular manipulated features,
whereas image fake news contents may contain multiple 1. Input: (1) : Regular Image
manipulated features. So, these features could not be ideal 2. Image Label (Real/Fake)
for effectively detecting image fake news. Previous studies
[21] on image fake news detection have utilized other 3. Output: Prediction Image as Fake or Real
models, which effectively identify fake images based on
4. Begin
general features but necessitate a significant number of
Convert Regular Image To Error Level Analysis Image
training data. ( )
CASIA 2.0 have been employed. The observed Accuracy Perform Error Level Analysis Image Resizer Method
rates were 83.2% and 94.7%, respectively. ( )
Xue et al.[21]built a model called “Multi-Vision Fusion 7. ficient et Extract Image Features Using EfficientNetB0
Neural Network” to detect pictures in fake news. To Framework ( )
extract the image features from an image in pixel domain, 8. Process Image Features Using Dense Layers
the visual modal module is utilized. Meanwhile, the ELA (
Framework ficient et )
is employed for feature extraction at the frequency
9. ( )
domain. The input image features are extracted in the ( )
10.Optimize Image Fake Detection Results
semantic detection phase using the pre-trained ResNet50. 11. end
The physical features module extracts the physical part to
recognize fake news images. All elements in the visual Algorithm1.The steps of the proposed Image Fake News Detection
model are connected and passed to PCA to reduce the
number of features. Then, the physical features module is
combined with the visual feature one in an ensemble 3-1- Pre-Processing Stage
module for fake image detection. The main preprocessing operations are:
To identify the final fake news image, XGBoost has been  The tiff image is removed. This is lossless
used. The datasets used in this approach are (D1) and image.
(D2), while the accuracy rates reach to 93.41% and
 Convert all the images into the RGB color
88.53%, respectively.
space.
However, to address these models' limitations, a new
 Resizing the image: in this step, all the ELA
approach to image fake news detection has been proposed
images convert to (128*128 pixels).
that uses the EfficientNetB0 model in this paper. The use
of EfficientNetB0 results in higher accuracy with fewer
parameters and shorter execution times, making it more 3-2- Error Level Analysis Method
efficient and faster option than other models.
After the preprocessing stage, all the images are
converted into ELA. ELA is a forensics technique. Created
44
Khalid Zamil1& Moghaddam Charkari, Image Fake News Detection Using Efficient NetB0 Model

by [23]to draw attention to the areas where a picture has

been compressed. It takes the feature of the lossy
compression method of manipulated images to identify
whether the image is tampered or not. Briefly, ELA is
done as follows;

| | (1)

Where the original image and reduce quality

image.
The difference between the two images is known as the
error levels related to the original pixels. The error level
indicates a number of changes that are directly connected
with the compression loss. If the variation is minimal, the
pixel has attained its local minimum for error at the
specified error rate. However, if there is substantial Fig. 2. a) The original image from Casia dataset and b) the image upon
alteration, the pixels are not in their local minimum and converting into ELA image.
may be extraneous[24],[25],[26]. In our proposed method,
the ELA algorithm is employed; accordingly, the re-saved 3-3- Image Feature Learning
version is compressed at a quality of 95%. Furthermore,
After the ELA images are produced, feature extraction is
the absolute variance between the quality and original
undertaken to establish feature vectors. In this paper, We
images is found. The dissimilarities among images
have employed the EfficientNetB0 model belongs to the
indicate the error levels associated with the initial pixels.
family EfficientNet from B0 to B7 in this study [27];
To improve the performance of the model, we fine-tune
Accordingly, each variant of EfficientNet introduces
the brightness of the images; a scale parameter has been different parameters as well as computational complexity
used to fine-tune the results; the value of this parameter
time. Comparing them, EfficientNetB0 finds the lowest
has been calculated as follow;
number of parameters while EfficientNetB7 needs the
highest number of parameters for training.
| (2) Since one of the urgent necessities in fake news detection
on social media is early and fast detection, we opted for
(3) EfficientNetB0, which uses fewest parameters. This allows
for faster execution time and reduces computational
where mix pixel is the maximum image pixel in . resources without sacrificing the quality of the results.
Figure 2 shows the original image after converting into EfficientNetB0 has been designed using a novel scaling
ELA image. Image a: method that optimizes the model's depth, width, and
resolution for the given computational resources. This
approach has resulted in a model that can achieve high
accuracy rate with fewer parameters, making it more
efficient and faster than other deep learning models like
ResNet, Inception, and VGG.
For feature extraction and added a
global_average_pooling2d layer to reduce the overall
number of parameters and hence limit overfitting. The
EfficientNetB0 is a trained model on the ImageNet
Dataset. To increase the model's efficiency, we re-trained
the EfficientNetB0 model on our dataset, and the first
layers from the EfficientNetB0 model pass to the
global_average_pooling2d layer to reduce overfitting. The
feature vector length ( ) is 1280 for the EfficientNetB0
model. Let
Image b: ( et ) (4)
 45
Journal of Information Systems and Telecommunication, Vol.12, No.1, January-March 2024

Where represents the activation function, i is the approximately 6% in accuracy and 5.2% in the case of F-
weight for each layer in EfficientNetB0, et is the score rates.
output from the layer. In future, we intend to expand our method to social media
datasets by extracting text from images and studying its
3-4- News Classification impact on fake news detection.

The final stage in our proposal predicts the image into 4-1- Parameters
two classes as fake or real. For image feature extraction,
we use the EfficientNetB0 model. Accordingly, the Transfer learning includes fine-tuning. We adjust our
sigmoid function is used to ascertain whether an image is model that has previously undergone training on the
authentic or fake. The Relu (Rectifier Linear Units) is used ImageNet dataset. As mentioned, the images are initially
in the dense layer as the activation function. We perform converted into ELA images and further resized to 128*128
the predictor for image fake news by the sigmoid function pixels. The EfficientNetB0 models have been used with
as: pre-trained ImageNet weights (just the part CNN feature
̃ ( ). (5) extraction, without prediction layers) by the
EfficientNetB0 model and re-train parts of the network on
Where indicates the parameters set of the sigmoid our dataset. ImageNet dataset was frozen so that the
function, and is the mapping function. weights of the ImageNet would not be affected by re-
Adam's optimizer has been utilized to optimize learning. training on our dataset. After training the network and
Binary Cross Entropy is applied to calculate the loss adjusting the parameters, we unfreeze the entire network.
function. ̃ ̃ ̃ hence ̃ denotes the probability of The last four layers (the top layer) related to the
the given image as actual (0). ̃ indicates the likelihood of classification process in CNN from the network are
the image being fake (1). removed and replaced with the proposed classification and
activation function layers.
4- Experimental Results & Analysis After re-training the network and extracting the features,
we added a GlobalAveragePooling2D Layer with a
dropout layer to eliminate the repetition in the features
We show the experiments conducted to assess the
resulting from the re-training process and overfitting.
effectiveness of the proposed model. This section details
To make it fit with our classes, we have added two dense
the dataset, outcomes, and comparison with other related
layers with a sigmoid function to predict whether the class
methods.
type is fake or real.
The learning rate 1e − 6 is set to warm up the FC. When
Fine-Tuning Given that social media has become a applying fine-tuning, we allow the warm-up stage to train
fundamental aspect of human daily life, detecting fake for 10 epochs based on our dataset. We will proceed to
news on these platforms has become a crucial issue. Those measure our network performance on the testing set after
methods used to spread fake news have evolved from text the warm-up phase. Table 1 describes the parameters used
to images and even videos. In this study, we proposed a to fine-tuning model.
method to detect fake images using the EfficientNetB0
model, a member of the CNN family that is trained on the Table 1. Hyper parameter settings for EfficientNetB0
ImageNet dataset. In general, images play a critical role in
news verification. In this regard, we have investigated on Hyperparameter Values
images to enhance fake news detection performance. The
ELA method is employed with the EfficientNetB0 model. Optimizer Adam
Furthermore, a global_average_pooling2d layer is added to Learning rate
reduce the number of parameters and to prevent No. of dense layers 2
overfitting. The EfficientNetB0 model has also been Dropout 0.5
trained on our dataset, and weights were set for the Batch Size 32
EfficientNetB0 during the training process. We validate Epochs 10
Total parameters 4,049,571
the effectiveness of feature learning on one of popular
Trainable Parameters 4,007,548
dataset, the CASAI. The proposed method achieves a
Non- Trainable Parameters 42,023
validation accuracy rate of 96.11%. The model is designed
to calculate the probability of the posts in the form of the
entered image being real or fake. The results outperformed
state-of-the-art methods on CASAI dataset, with a rate of
 46
Khalid Zamil1& Moghaddam Charkari, Image Fake News Detection Using Efficient NetB0 Model

4-2- Experimental Setup Recall, and precision. Accuracy indicates how well the
model classifies the images as real or fake.
The model was produced using a machine on Colab, The F-score measures the consistent mean of Precision and
employing the Keras library and the Google TensorFlow Recall; the performance of the proposed model, as shown in Fig
3, is found by four measures, Accuracy, Precision, Recall, and F-
frame. To choose optimal hyperparameters, we have studied
Score denoted as follows;
different batch sizes and dropout probabilities. The best
results were achieved by utilizing the Adam optimizer with a
learning rate of a batch size of 32, and training for 20 ( )
epochs. The hyperparameter values in this study are shown
( )
in table 2. Each experiment has been carried out randomly.
Accordingly, the CASIA dataset is split into 80% as training ( )
and 20% as validation. The final findings were obtained ( )
when the ultimate level of accuracy was attained. ( )
( )
Where,
Table 2 . Hyperparameter values in the proposed model
( )
( )
Hyperparameter Values
( )
( )
Optimizer Adam
Learning rate AUC represented a level of separability. It indicates the
Dropout 0.5 model's capability to distinguish between classes.
Batch Size 32 Although AUC has not been taken into comparison with
Epochs 20 other methods, it is significant for checking in
Total params 168,129 classification problems; Fig 4 describes the confusion
Trainable Params 168,129 matrix of the proposed method on the CASIA data set. Fig
Non- Trainable Parameters 0 5 represents an ROC graph used to evaluate an imbalanced
dataset, which is essential in binary classification. Table 4
4-3- CASIA 2.0 Dataset displays a comparison between our proposed approach and
other baseline methods. Refer to Table 4, [20]used the
There are 12,616 images in the CASIA 2.0 dataset, high pass filter with CNN to detect fake images. [29]Used
where 5124 of them are manipulated, and the remaining the VGG19 model to detect fake images. MVNN[30]
7492 images are legitimate. Copy-move and image- employed physical and semantic visual features to find
splicing techniques are used to manipulate the images. fake news. In [31] utilized a CNN to extract features that
While performing tampering to the image, cropping and help in the identification of fake news. As shown in Table
resizing are also done [28]. The number of CASIA images 4, the proposed model can efficiently capture the modified
is shown in table 3. characteristics in the fake image.

Performance metrics of the proposed model

Table 3 The statistics of CASIA V.2 dataset
99.00% 98.13%
98.00% 97.42%
Image type Image size 97.00%
96.83%
96.11%
96.00%
Authentic – image 7200 95.00%
Tamper – image 5123 94.00%
93.00%
92.00%
4-4- Experimental Results 91.00%
90.00%
CASIA
As mentioned in the previous section, the EfficientNetB0
Accuracy Precision Recall F-score
model has been employed to learn the essential features,
represented by the fewest number of parameters with high Fig. 3. The results of the proposed model's performance
efficiency compared to the EfficientNetB1 model to
EfficientNetB7. We have conducted our experiments on
the CASIA dataset that contains images. The highest
accuracy rate has been found as 96.11% in value. Four
assessment measures have been used to evaluate the
experimental results. Those are F1-score, Accuracy,
 47
Journal of Information Systems and Telecommunication, Vol.12, No.1, January-March 2024

Table 3 . Comparison of different models with our model employed with the EfficientNetB0 model. Furthermore, a
global_average_pooling2d layer is added to reduce the
Method Year Accuracy F-score number of parameters and to prevent overfitting. The
CNN [20] 2021 94.7% 95% EfficientNetB0 model has also been trained on our dataset,
VGG19 [29] 2019 74.07% 79.11% and weights were set for the EfficientNetB0 during the
MVNN[30] 2019 89.12 94.53 training process. We have validated the effectiveness of
CNN [31] 2017 74% 74.4% feature learning on one of popular dataset, the CASAI. The
Our proposed 2023 96.11% 97.42 % proposed method achieves a validation accuracy rate of
EfficientNetB0 model 96.11%. The model is designed to calculate the probability
of the posts in the form of the entered image being real or
fake. The results outperformed state-of-the-art methods on
CASAI dataset, with a rate of 6% in accuracy and 5.2%
in the case of F-score rates.
In future, we intend to extend our method to social media
datasets by extracting text from images and studying its
impact on fake news detection. Furthermore, introducing
an explanatory model is another further direction of our
research.
References
[1] M. Celliers and M. Hattingh, “A Systematic Review on Fake
News Themes Reported in Literature,” in Lecture Notes in
Computer Science (including subseries Lecture Notes in
Artificial Intelligence and Lecture Notes in Bioinformatics),
Springer, 2020, pp. 223–234. doi: 10.1007/978-3-030-45002-
1_19.
[2] X. Zhang and A. A. Ghorbani, “An overview of online fake
news: Characterization, detection, and discussion,” Inf
Fig. 4. Confusion matrix of the proposed method on CASIA 2.0 dataset Process Manag, vol. 57, no. 2, p. 102025, 2020, doi:
10.1016/j.ipm.2019.03.004.
[3] W. S. Paka, R. Bansal, A. Kaushik, S. Sengupta, and T.
Chakraborty, “Cross-SEAN: A cross-stitch semi-supervised
neural attention model for COVID-19 fake news detection,”
Appl Soft Comput, vol. 107, p. 107393, 2021, doi:
10.1016/j.asoc.2021.107393.
[4] Y. Wang et al., “Weak supervision for fake news detection
via reinforcement learning,” AAAI 2020 - 34th AAAI
Conference on Artificial Intelligence, no. December 2019,
pp. 516–523, 2020, doi: 10.1609/aaai.v34i01.5389.
[5] N. Guimarães, Á. Figueira, and L. Torgo, “An organized
review of key factors for fake news detection,” pp. 1–10,
2021, [Online]. Available: http://arxiv.org/abs/2102.13433
[6] S. Preston, A. Anderson, D. J. Robertson, M. P. Shephard,
and N. Huhe, “Detecting fake news on Facebook: The role of
emotional intelligence,” PLoS One, vol. 16, no. 3 March, pp.
Fig. 5. AUROC curve on the CASIA 2.0 dataset 1–13, 2021, doi: 10.1371/journal.pone.0246757.
[7] P. Meel and D. K. Vishwakarma, “Fake news, rumor,
5- Conclusion information pollution in social media and web: A
contemporary survey of state-of-the-arts, challenges and
Given that social media has become a fundamental opportunities,” Expert Syst Appl, vol. 153, p. 112986, 2020,
aspect of human daily life, detecting fake news on these doi: 10.1016/j.eswa.2019.112986.
platforms has become a crucial issue. Those methods used [8] H. Allcott and M. Gentzkow, “Social media and fake news in
to spread fake news have evolved from text to images and the 2016 election,” Journal of Economic Perspectives, vol.
even videos. In this study, we proposed a method to detect 31, no. 2, pp. 211–236, 2017, doi: 10.1257/jep.31.2.211.
fake images using the EfficientNetB0 model, a member of [9] S. Raza and C. Ding, “Fake news detection based on news
content and social contexts: a transformer-based approach,”
the CNN family that is trained on the ImageNet dataset. In
Int J Data Sci Anal, vol. 13, no. 4, pp. 335–362, May 2022,
general, images play a critical role in news verification. In doi: 10.1007/s41060-021-00302-z.
this regard, we have investigated on images to enhance [10] X. Zhang and A. A. Ghorbani, “An overview of online
fake news detection performance. The ELA method is fake news: Characterization, detection, and discussion,” Inf
 48
Khalid Zamil1& Moghaddam Charkari, Image Fake News Detection Using Efficient NetB0 Model

Process Manag, vol. 57, no. 2, p. 102025, 2020, doi: [25] Paganini and Pierluigi, "Photo forensics: Detect
10.1016/j.ipm.2019.03.004. Photoshop manipulation with error level analysis." Chief
[11] N. Hoy and T. Koulouri, “A Systematic Review on the Information Security Officer at Bit4Id, 2013.
Detection of Fake News Articles,” Oct. 2021, [Online]. [26] H. Farid, “Exposing Digital Forgeries from JPEG
Available: http://arxiv.org/abs/2110.11240 Ghosts,” IEEE transactions on information forensics and
[12] J. Jing, H. Wu, J. Sun, X. Fang, and H. Zhang, security 154-160, 4.1 .2009.
“Multimodal fake news detection via progressive [27] M. Tan and Q. V. Le, “EfficientNet: Rethinking model
scaling for convolutional neural networks,” 36th
fusion networks,” Inf Process Manag, vol. 60, no. 1, International Conference on Machine Learning, ICML 2019,
Jan. 2023, doi: 10.1016/j.ipm.2022.103120. vol. 2019-June, pp. 10691–10700, 2019.
[13] A. Biswas, D. Bhattacharya, K. Anil Kumar, and [28] W. and T. T. Jing Dong, “CASIA IMAGE
A. Professor, “DeepFake Detection using 3D-Xception TAMPERING DETECTION EVALUATION DATABASE
Net with Discrete Fourier Transformation,” Journal of Jing Dong , Wei Wang and Tieniu Tan Institute of
Information Systems and Telecommunication (JIST) 3, no. Automation , Chinese Academy of Sciences,” pp. 422–426,
35. 2021. 161. 2013.
[14] S. Hangloo and B. Arora, “Combating multimodal fake [29] D. Khattar, M. Gupta, J. S. Goud, and V. Varma,
news on social media: methods, datasets, and future “MvaE: Multimodal variational autoencoder for fake news
perspective,” Multimed Syst, vol. 28, no. 6, pp. 2391–2422, detection,” The Web Conference 2019 - Proceedings of the
Dec. 2022, doi: 10.1007/s00530-022-00966-y. World Wide Web Conference, WWW 2019, pp. 2915–2921,
[15] D. K. Vishwakarma, D. Varshney, and A. Yadav, 2019, doi: 10.1145/3308558.3313552.
“Detection and veracity analysis of fake news via scrapping [30] P. Qi, J. Cao, T. Yang, J. Guo, and J. Li, “Exploiting
and authenticating the web search,” Cogn Syst Res, vol. 58, multi-domain visual information for fake news detection,”
pp. 217–229, Dec. 2019, doi: 10.1016/j.cogsys.2019.07.004. Proceedings - IEEE International Conference on Data
[16] D. Mangal and Di. K. Sharma, “Fake News Detection Mining, ICDM, vol. 2019-Novem, no. Icdm, pp. 518–527,
with Integration of Embedded Text Cues and Image 2019, doi: 10.1109/ICDM.2019.00062.
Features,” ICRITO 2020 - IEEE 8th International Conference [31] F. Yu, Q. Liu, S. Wu, L. Wang, and T. Tan, “A
on Reliability, Infocom Technologies and Optimization Convolutional Approach for Misinformation Identification,”
(Trends and Future Directions), pp. 68–72, 2020, doi: 2017. [Online]. Available:
10.1109/ICRITO48877.2020.9197817. http://www.npr.org/2016/11/08/500686320/did-social-media-
[17] A. Mahmoodzadeh, “Human Activity Recognition
based on Deep Belief Network Classifier and Combination of
Local and Global Features,” .” J. Inf. Syst. Telecommun 9,
2021.
[18] Z. Jin, J. Cao, Y. Zhang, J. Zhou, and Q. Tian, “Novel
Visual and Statistical Image Features for Microblogs News
Verification,” IEEE Trans Multimedia, vol. 19, no. 3, pp.
598–608, 2017, doi: 10.1109/TMM.2016.2617078.
[19] F. Marra, D. Gragnaniello, D. Cozzolino, and L.
Verdoliva, “Detection of GAN-Generated Fake Images over
Social Networks,” Proceedings - IEEE 1st Conference on
Multimedia Information Processing and Retrieval, MIPR
2018, pp. 384–389, 2018, doi: 10.1109/MIPR.2018.00084.
[20] B. Singh and D. K. Sharma, “SiteForge: Detecting and
localizing forged images on microblogging platforms using
deep convolutional neural network,” Comput Ind Eng, vol.
162, no. October, 2021, doi: 10.1016/j.cie.2021.107733.
[21] J. Xue, Y. Wang, S. Xu, L. Shi, L. Wei, and H. Song,
MVFNN: Multi-Vision Fusion Neural Network for Fake
News Picture Detection, vol. 1300, no. 2018. Springer
International Publishing, 2020. doi: 10.1007/978-3-030-
63426-1_12.
[22] C. Boididou et al., “Verifying multimedia use at
MediaEval 2015,” CEUR Workshop Proc, vol. 1436, no.
September, 2015.
[23] I. B. K. Sudiatmika, F. Rahman, Trisno, and Suyoto,
“Image forgery detection using error level analysis and deep
learning,” Telkomnika (Telecommunication Computing
Electronics and Control), vol. 17, no. 2, pp. 653–659, 2019,
doi: 10.12928/TELKOMNIKA.V17I2.8976.
[24] N. Krawetz, “A Picture’s Worth... Digital Image
Analysis and Forensics,” 2007. [Online]. Available:
www.hackerfactor.com

View publication stats