2023 International Conference on Control, Communication and Computing (ICCC) 19-21 May 2023

Deep fake Detection using deep learning

techniques: A Literature Review
2023 International Conference on Control, Communication and Computing (ICCC) | 979-8-3503-3412-8/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICCC57789.2023.10164881

Amala Mary∗ , Anitha Edison†

† Computer Vision Lab, College of Engineering Trivandrum, Kerala.
† Affiliated to APJ Abdul Kalam Technological University, Trivandrum, Kerala, India.
Email: ∗ [email protected], † [email protected]

Abstract—Deep learning is a sophisticated and adaptable models are used to train on a dataset and create fictitious
technique that has found widespread use in fields such as natural videos and pictures. In reality, the widespread availability of
language processing, machine learning, and computer vision. videos/images on social media might aid people in creating
It is one of the most recent deep learning-powered applica-
tions to emerge. Deep fakes are altered, high-quality, realistic plausible rumors and false information that could lead to
videos/images that have lately gained popularity. Many incredible creating a negative impact on society.
uses of this technology are being investigated. Malicious uses of According to recent research, deep fake videos and im-
fake videos, such as fake news, celebrity pornographic videos, ages are widely disseminated on social media. Deep fake
financial scams, and revenge porn are currently on the rise in video/image detection has therefore become increasingly cru-
the digital world. As a result, celebrities, politicians, and other
well-known persons are particularly vulnerable to the Deep fake cial and important. Many deep learning approaches such as
detection challenge. Numerous research has been undertaken in Recurrent Neural Network (RNN) [2], Convolutional Neural
recent years to understand how deep fakes function and many Network (CNN) [3], and Long short-term memory (LSTM)
deep learning-based algorithms to detect deep fake videos or [4], [5] are proposed to detect deep fake videos/images. And
pictures have been presented. this will bring up more research on this area.
This study comprehensively evaluates deep fake production and
detection technologies based on several deep learning algorithms. This study focuses on deep fake detection algorithms that
In addition, the limits of current approaches and the availability have previously been deployed. It primarily covers classic
of databases in society will be discussed. A deep fake detection detection methods as well as deep Learning based methods
system that is both precise and automatic. Given the ease with such as CNN, RNN, and LSTM. The first section of the study
which deep fake videos/images may be generated and shared, provides a quick overview of deep fakes and their societal
the lack of an effective deep fake detection system creates a
serious problem for the world. However, there have been various consequences. The evaluated overview of relevant studies is
attempts to address this issue, and deep learning-related solutions mentioned in Section II. Then Section III discusses several
outperform traditional approaches. detection approaches and strategies, with an emphasis on
Index Terms—-Deep Fakes, Deep Learning, Fake Generation, conventional deep learning. Section IV provides an overview
Fake Detection, Machine Learning of recently published datasets. And Section V contains the
conclusion part.
I. I NTRODUCTION
II. R ELATED W ORKS
The face is the most unique feature of human beings.
A. Deep Learning
With the rapid advancement of face synthesis technology, the
security risk provided by face alteration is becoming increas- It is a machine learning approach similar to neural networks
ingly significant. Deep fake is one of the artificial intelligence [6] [7] and refers to the use of several hidden units in a net-
technology in which one person’s face is superimposed on top work. Its basic architecture, influenced by artificial networks,
of another person’s face without his/her permission. employs an unlimited number of hidden units of bound size.
Deep learning is a powerful and valuable technology that This is done to evoke additional information from the input
has been applied in many fields, including machine learn- data. The complexity of the trained data determines the number
ing, computer vision, and natural language processing. As a of hidden layers [6]. More complicated data need more hidden
result of advancements in deep learning, modifying digital layers to generate accurate findings. In recent years, it has been
material, and creating synthetic content has become quite successfully employed in a wide range of domains and it will
simple. Generative adversarial networks (GANs) [1] and deep continue to be used.
learning algorithms are used to create fake images and videos 1) Convolutional Neural Network: CNN is the deep neural
that are difficult for humans to distinguish from the real network architecture most commonly used. It has an input
ones. These are produced using enormous datasets, then those layer, an output layer, and one or more hidden layers, just
like other neural networks. In CNN [3], the hidden layers
first read the inputs from the first layer and then execute a

979-8-3503-3412-8/23/$31.00 ©2023 IEEE

convolution mathematical operation on the input values. In
addition to matrix multiplication, CNN employs non-linearity
activation methods such as Rectified Linear Units (RELU)
and extra convolutional approaches such as pooling layers. To
minimize the complexity of the data, pooling layers provide
outputs using methods such as average pooling.
2) Recurrent Neural Network: It is another artificial neu-
ral network application that can learn characteristics from
sequence data [2]. Basically, RNN is built from a variety Fig. 1. Basic architecture of GAN [1]
of hidden layers, each with its own bias and weight. The
connection between the nodes in an RNN-based direct cycle
graph runs sequentially. By offering a recurrent hidden state swap faces in movies [9] [10]. Similar to GANs, Fake App
that encapsulates time-scale dependencies, it can handle a employs an autoencoder to extract latent features from images
temporal sequence. of human faces and a decoder to re-extract those features
3) Long Short-Term Memory: It is a sort of artificial RNN from the same images. This method is effective because it can
that manages long-term dependencies [4] [5]. The full data create fake videos that are surprisingly real and challenging to
sequence may be learned using the feedback connections in distinguish apart from the real thing. Another well-known deep
LSTM. The input gate, forget gate and output gate make up fake method is based on a generative adversarial network and
the basic LSTM architecture. The cell state remembers the that is called VGGFace (GAN). The structure of VGGFace
values from prior intervals and stores them. The input gate first [11] is enhanced by the addition of two layers known as
selects the values that ought to be written into the cell state. adversarial loss and perceptual loss. In order to capture hidden
The forget gate may logically select which information has to features of facial images these layers are added to the auto
be forgotten by employing a sigmoid function. In which the encoder-decoder. This produces more convincing and realistic
information from the present moment should be taken into fake images.
account in the following phase is decided by the output gate.
B. Deep fake Detection
III. D EEP FAKE G ENERATION AND D ETECTION Deep learning has shown considerable achievement in the
The deep fake approach creates fake videos/images using identification of deep fakes. In the section below first go
GANs techniques. In this part, first, provide a summary of through the deep learning-based image detection models then
the current software and resources for producing deep-fake continue on to the video detection models.
images/videos. And then go through different deep-learning
detection methods to address this problem. 1) Image Detection Methods: Several techniques have been
explored to recognize GAN-generated pictures using deep
A. Deep fake Generation networks. Tariq et al. [12] proposed using neural networks to
GANs are deep neural network that is frequently used to detect false GAN videos. This methodology analyzes image
create deep fakes. GANs have the benefit of learning from statistical components and enhances recognition of artificially
a group of training data sets and producing a sample of produced fake facial photos. Nhu et al. [13] contribute
information with the same characteristics and qualities. For another strategy for identifying fraudulent images generated
example, GANs might be used to substitute a “genuine” by GANs that is based on a deep CNN. This technique starts
picture or video of a person with an “altered one” [8]. GAN by utilizing a deep learning network to evoke facial attributes
architecture consists of two neural network components: an from face identification networks.
encoder and a decoder. The model uses the encoder to train Xuan et al. [14] employed image preprocessing techniques,
on a vast data set in order to generate fictitious data. The such as Gaussian blur and Gaussian noise. This improves
decoder is then used to discriminate between real and fake the mathematical similarity between authentic photos and
data. This model requires a significant amount of input in imitations at the pixel level, allowing the scientific classifier to
order to create realistic-looking faces (images and videos). The pick up more intrinsic features, and improves generalization
GAN architecture is depicted in Figure 1. To produce a fake capacity than earlier techniques for picture forensics [15]
sample, the encoder is first fed random input seeds, as seen in [16].
the image. The decoder is trained using these fake samples. Zhao et al. [17] recently introduced a methodology for deep
This decoder is a binary classifier that takes in both real and fake detection utilizing the self-consistency of local source
fake data and then uses a Softmax function to differentiate features, which are spatially-local, content-independent details
between them. of pictures. A CNN model employs a unique representation
Numerous deep fake applications have been used for a long learning approach to extract these source features, which are
time. The first technique that has been widely utilized for represented as down-sampled feature maps referred to as
deep fake production is Fake App. This app uses an auto pairwise self-consistency learning. This aims to punish feature
encoder-decoder pairing structure created by a Reddit user to vector pairings that correspond to areas in the same picture
with poor cosine similarity scores. When dealing with false with around 47.7% of the photos being male and 52.3% being
pictures created by technologies that output the entire image female.
directly and whose source features are constant throughout
each point inside each image, it could have a disadvantage. B. VGGFace2
The large-scale face dataset is known as VGGFace2 consists
2) Video Detection Methods: Due to the huge loss of frame of 3 million face photographs of nine thousand unique indi-
content during video compression, existing deep learning viduals, with an average of more than 300 photos per subject.
algorithms for image identification cannot effectively detect The Google search engine provided images, which have a lot
bogus videos. of information like age, race, lighting, and occupation.
The severe deterioration of the frame data following video
compression prevents the majority of image recognition tech- C. Flickr-Faces-HQ (FFHQ)
niques from being employed for videos [18]. Additionally, This database contains human face information. And the
videos provide a problem for techniques intended to identify database FFHQ contains 70,000 face images with a high-
only still fake images since their temporal features vary across quality resolution created by GAN. The author claims that
sets of frames. Based on the discovery that temporal coherence the dataset underwent pre-processing to reduce the size of the
is not properly preserved in the synthesis process of deep collection and remove picture noise.
fakes, Sabir et al. [19] used spatiotemporal characteristics of
video streams to detect deep fakes. Frame-by-frame editing D. 100K-Faces
is used in video editing. A framework on which low-level A well-known publicly accessible dataset called 100K-
face manipulation defects are expected to further appear as Faces contains 100,000 original human photos created with
temporal distortions with irregularities between the frames. StyleGAN. StyleGAN was used to create photographs with a
However, deep learning algorithms frequently employ face flat backdrop from a big dataset of more than 29,000 images
photos from the internet that typically display people with collected from 69 distinct models.
wide eyes; fewer pictures of persons with closed eyes may
be seen online. As a result, deep fake algorithms are unable V. C OMMONLY USED E VALUATION PARAMETERS
to generate fake faces that blink often in the absence of The effectiveness of various deep fake detection approaches
photographs of actual people doing so. Deep fakes, in other is evaluated using the metrics listed below. They are:
words, have far lower blink rates than regular videos. 1) Accuracy:
Li et al. [20] trim eye regions from the films and distribute The most basic metric for assessing a classification
them to long-term recurrent convolutional networks (LRCN) model’s performance is accuracy. According to the given
[21] for dynamic state prediction in order to distinguish equation, classification accuracy is calculated by dividing
between authentic and fake videos. the number of true results by the total number of results.
A deep learning technique that is used to detect deep fakes was
presented in [22]. The UADFV and Deep fakeTIMIT deep fake Tp + Tn
datasets are used to assess the proposed approach. The total Accuracy = × 100 (1)
Tp + Tn + Fp + Fn
number of frames in the UADFV database [23] is 32,752.
It consists of 49 authentic videos and 49 fraudulent videos. where Tp , Tn , Fp , and Fn represent the true positive, true
The suggested strategy eliminates the necessity for deep fake negative, false positive, and false negative rates.
videos to be created as negative examples before training
the detection algorithms. Instead, the negative instances are 2) Area Under Curve:
created dynamically by deleting the face region from the It gives a total performance evaluation across all potential
original picture, applying Gaussian blur to a scaled image classification thresholds.
of a random choice, and then stretching back to the original
image after numerous scale alignments. Compared to previous 3) Precision:
approaches that call for the creation of deep fakes in advance, The following equation, which reflects the ratio of the
this requires a significant reduction in time and computing total number of positive items divided by the number of
resources. true positives, is used to calculate accuracy during the
classification step.
IV. D IFFERENT TYPES OF DATASETS
Tp
Next, we discuss the databases used for the identification P recision = × 100 (2)
of deep fakes. Tp + Fn

A. Fake Face Dataset (DFFD) 4) Receiver Operating Characteristics(ROC):

It includes 100,000 and 200,000 false photos that taken The recall values are represented on the y-axis, while the
from ProGAN and StyleGAN models. The majority of the specificity values are plotted on the x-axis is called ROC.
samples in the collection are between the ages of 21 and 50,
 TABLE I
OVERVIEW O F P OPULAR D EEP FAKE DETECTION M ETHODS

SL.No. Authors Methodology Techniques Key Features Databases Used

1 Li, Yuezun, Ming-Ching Eye blinking Long term recurrent CNN Use LRCN to understand Consist of 49 interview
Chang, and Siwei Lyu. the temporal patterns of & presentation videos, &
[20] eye blinking. their corresponding
generated deepfakes.
2 Afchar, Darius et al. [18] MesoNet CNN Two deep networks Two databases: Deep fake
Mes0-4 & one constituted from
Mesoinception-4 are online videos & Face
introduced to examine Forensics one created by
deep false videos at the face2face approach
mesoscopic level of
analysis.
3 Sabir, Ekraam et al. [19] Spatio-temporal features RCN RCN, which combines FaceForensics++ dataset,
with RCN the convolutional network including 1000 videos
DenseNet with the gated
recurrent unit cells, is
used to investigate
temporal differences
between frames.
4 Chintha, Akash et al. [9] Spatio-temporal features convolutional For face feature FaceForensics++,
with LSTM bidirectional recurrent extraction, an Celeb-DF & ASVSpoof
LSTM network XceptionNet CNN is 2019 logical Access
employed, and audio audio dataset
embeddings are generated
by stacking numerous
convolution modules.
5 Fernandes, Steven et al. Using attribution based ResNet50 model Without access to training VidTIMIT, COHFACE
[24] confidence (ABC) metric pretrained on VGGFace2 data, deep fake videos are
detected using the ABC
measure.
6 Ciftci, Umur Aybars, Ilke FakeCatcher CNN Biological signals are not UADFV, FaceForensics,
Demir, and Lijun Yin. spatially and temporally FaceForensics++,
[25] well maintained in deep Celeb-DF
fakes, they are extracted
from portrait videos and
used as an implicit
descriptor of authenticity.
7 Yang, Xin, Yuezun Li, Head poses Support Vector Machine Features are evoked using UADFV is made up of
and Siwei Lyu. [23] 68 landmarks of the face 49 false videos and their
region. actual videos, as well as
241 real photos & 252
fake images from the
DARPA MediFor GAN
Video/image Challenge.
8 Xuan, Xinsheng et al. Preprocessing combined DCGAN, WGAN-GP & Improve deep learning Real dataset: CelebA-HQ,
[14] with deep network PGGAN. models’ generalisation Fake datasets: generated
capacity to recognize by DCGAN, WGANGP
GAN produced images. and PGGAN
9 Güera, David, and Passed through CNN for CNN A model which HOHA dataset & deep
Edward J. Delp. [26] feature extraction automatically detect deep fake videos from multiple
fake videos using hosting websites
recurrent neural network
10 Wang, Sheng-Yu et al. Using common artifacts ResNet-50 pre-trained Using a large number of A new dataset of
[27] of CNN-generated images with ImageNet fictitious pictures CNN-generated images,
produced by a namely ForenSynths,
high-performing consisting of synthesized
unconditional GAN images from 11 models
model, i.e., PGGAN, such as StyleGAN,
train the classifier and super-resolution methods
assess how well it & FaceForensics++
generalizes to other
CNN-generated images.
11 Gandhi, Apurva, and Defenses against VGG & ResNet Introduce adversarial 5,000 real images from
Shomik Jain [28] adversarial perturbations perturbations to enhance CelebA & 5,000 fake
in deep fakes deep fakes and fool deep images generated by the
fake detectors. “Few-Shot Face
Translation GAN”
method
12 Li, Lingzhi et al. [29] Face X-ray CNN Instead of collecting the FaceForensics++,
synthesized artefacts of DeepfakeDetection
certain operations, try to (DFD), DFDC &
find the border between Celeb-DF
the target and original
faces.
SL.No. Authors Methods Classifiers/Techniques Key Features Datasets Used
13 Zhang, Ying, Lilei Bag of words & shallow SVM, RF, MLP Using the bag of words The popular LFW face
Zheng, and Vrizlynn LL classifiers approach, extract database, which has
Thing. [30] discriminant 13,223 photos at a
characteristics and input resolution of 250x250.
these into SVM, RF, and
MLP for binary
classification: real versus
fake.
14 Chen, Zehao, and Hua Manipulted Face detector CNN Based on the Multilevel FF, Celeb-A, FF++
Yang. [31] Facial Conceptual
Extraction and Cascade
Attention Mechanism,
face identification
methodology is modified.
15 Durall, Ricard et al. [32] Unmasking Deepfakes SVM, LR, k-MN Method is based on a Celeb-A, FF++
traditional frequency
domain analysis, which is
followed by a basic
classifier.
16 Li, Xiaodan et al. [33] Sharp Multi-Instance MIL In contrast to the classic Celeb-DF, FF, FF+,
Learning MIL, which creates a DFDC
straight mapping from
instance embeddings to
bag prediction before
moving on to instance
prediction, a sharp MIL
(S-MIL) is presented.
17 Rana, Md Shohel, and DeepFake Stack CNN For identifying such Celeb-DF, FF++
Andrew H. Sung. [34] altered videos, use the
DeepfakeStack deep
ensemble learning
approach.
18 Do, Nhu-Tai, In-Seop Na, Forensics Face detection CNN A deep CNN to detect Celeb-A
and Soo-Hyung Kim. [13] forensics face.
19 Nguyen, Huy H et al. Multi-task Learning CNN CNN that use the FF, FF++
[35] multi-task learning
strategy to find the altered
areas for each query
while also simultaneously
detecting altered pictures
and videos.
20 Guo, Zhiqing et al. [36] Using deep features A new CNN model, High-level forensics Glow model was used to
extracted by CNN namely SC net characteristics may be produce a collection of
automatically learned by 321,378 face photos from
CNN-based SCnet from the CelebA face image
visual input due to a dataset.
hierarchical feature
extraction block
constructed by stacking
four convolutional layers.

5) F1-score: deep learning-based approaches have recently been put out to

It represents the arithmetic mean of accuracy and recall. deal with this problem and effectively identify fake images and
P recision × Recall videos. The first section discussed the existing programs and
F1 = 2 × × 100 (3) technologies that are extensively used to make fake photos and
P recision + Recall
videos. And in the second section discuss the different type of
techniques that are used for deep fake images and videos. Also,
provide details of available datasets and evaluation metrics that
VI. C ONCLUSION are used for deep fake detection.
Various researchers have created a number of deep-learning Despite the fact that deep learning has done well in detecting
approaches for deep fake images and videos. Due to the deep fakes, the quality of deep fakes has been increasing.
extensive availability of photographs and videos in social In order to recognize fake videos & photos properly must
media material, deep fakes had grown in popularity. This is be enhanced current deep learning approaches. Furthermore,
especially crucial in social networking sites that make it simple given present deep learning approaches, it is unknown how
for users to spread and share such fake information. Numerous to identify the number of layers necessary and the appropriate
architecture for deep fake detection. To improve their capacity [22] Y. Li and S. Lyu, “Exposing deepfake videos by detecting face warping
to cope with the ubiquitous impacts of deep fakes and mitigate artifacts,” arXiv preprint arXiv:1811.00656, 2018.
[23] X. Yang, Y. Li, and S. Lyu, “Exposing deep fakes using inconsistent
their consequences, social media companies are integrating head poses,” in ICASSP 2019-2019 IEEE International Conference on
deep fake detection tools. Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2019, pp.
8261–8265.
R EFERENCES [24] S. Fernandes, S. Raj, R. Ewetz, J. S. Pannu, S. K. Jha, E. Ortiz,
[1] M. Mirza and S. Osindero, “Conditional generative adversarial nets,” I. Vintila, and M. Salter, “Detecting deepfake videos using attribution-
arXiv preprint arXiv:1411.1784, 2014. based confidence metric,” in Proceedings of the IEEE/CVF Conference
[2] Y. Bengio, P. Simard, and P. Frasconi, “Long short-term memory,” IEEE on Computer Vision and Pattern Recognition Workshops, 2020, pp. 308–
Trans. Neural Netw, vol. 5, pp. 157–166, 1994. 309.
[3] I. Goodfellow, Y. Bengio, and A. Courville, Deep learning. MIT press, [25] U. A. Ciftci, I. Demir, and L. Yin, “Fakecatcher: Detection of synthetic
2016. portrait videos using biological signals,” IEEE transactions on pattern
[4] S. Hochreiter, “Ja1 4 rgen schmidhuber (1997).“long short-term mem- analysis and machine intelligence, 2020.
ory”,” Neural Computation, vol. 9, no. 8. [26] D. Güera and E. J. Delp, “Deepfake video detection using recurrent
[5] M. Schuster and K. Paliwal, “Networks bidirectional reccurent neural,” neural networks,” in 2018 15th IEEE international conference on
IEEE Trans Signal Proces, vol. 45, pp. 2673–2681, 1997. advanced video and signal based surveillance (AVSS). IEEE, 2018,
[6] J. Hopfield et al., “Rigorous bounds on the storage capacity of the dilute pp. 1–6.
hopfield model,” Proceedings of the National Academy of Sciences, [27] S.-Y. Wang, O. Wang, R. Zhang, A. Owens, and A. A. Efros, “Cnn-
vol. 79, pp. 2554–2558, 1982. generated images are surprisingly easy to spot... for now,” in Proceedings
[7] Y. Wu, M. Schuster, Z. Chen, Q. V. Le, M. Norouzi, W. Macherey, of the IEEE/CVF conference on computer vision and pattern recognition,
M. Krikun, Y. Cao, Q. Gao, K. Macherey, et al., “Google’s neural 2020, pp. 8695–8704.
machine translation system: Bridging the gap between human and [28] A. Gandhi and S. Jain, “Adversarial perturbations fool deepfake de-
machine translation,” arXiv preprint arXiv:1609.08144, 2016. tectors,” in 2020 International joint conference on neural networks
[8] L. Nataraj, T. M. Mohammed, B. Manjunath, S. Chandrasekaran, (IJCNN). IEEE, 2020, pp. 1–8.
A. Flenner, J. H. Bappy, and A. K. Roy-Chowdhury, “Detecting gan [29] L. Li, J. Bao, T. Zhang, H. Yang, D. Chen, F. Wen, and B. Guo, “Face
generated fake images using co-occurrence matrices,” Electronic Imag- x-ray for more general face forgery detection,” in Proceedings of the
ing, vol. 2019, no. 5, pp. 532–1, 2019. IEEE/CVF conference on computer vision and pattern recognition, 2020,
[9] B. Zi, M. Chang, J. Chen, X. Ma, and Y.-G. Jiang, “Wilddeepfake: A pp. 5001–5010.
challenging real-world dataset for deepfake detection,” in Proceedings of [30] H. Cui, A. Liu, X. Zhang, X. Chen, K. Wang, and X. Chen, “EEG-
the 28th ACM international conference on multimedia, 2020, pp. 2382– based emotion recognition using an end-to-end regional-asymmetric
2390. convolutional neural network,” Knowledge-Based Systems, vol. 205, p.
[10] H. A. Khalil and S. A. Maged, “Deepfakes creation and detection using 106243, 2020.
deep learning,” in 2021 International Mobile, Intelligent, and Ubiquitous [31] Z. Chen and H. Yang, “Attentive semantic exploring for manipulated
Computing Conference (MIUCC). IEEE, 2021, pp. 1–4. face detection,” in ICASSP 2021-2021 IEEE International Conference
[11] J. Luttrell, Z. Zhou, Y. Zhang, C. Zhang, P. Gong, B. Yang, and R. Li, on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021,
“A deep transfer learning approach to fine-tuning facial recognition pp. 1985–1989.
models,” in 2018 13th IEEE Conference on Industrial Electronics and [32] R. Durall, M. Keuper, F.-J. Pfreundt, and J. Keuper, “Unmasking
Applications (ICIEA). IEEE, 2018, pp. 2671–2676. deepfakes with simple features,” arXiv preprint arXiv:1911.00686, 2019.
[12] S. Tariq, S. Lee, H. Kim, Y. Shin, and S. S. Woo, “Detecting both ma- [33] X. Li, Y. Lang, Y. Chen, X. Mao, Y. He, S. Wang, H. Xue, and
chine and human created fake face images in the wild,” in Proceedings Q. Lu, “Sharp multiple instance learning for deepfake video detection,”
of the 2nd international workshop on multimedia privacy and security, in Proceedings of the 28th ACM international conference on multimedia,
2018, pp. 81–87. 2020, pp. 1864–1872.
[13] N.-T. Do, I.-S. Na, and S.-H. Kim, “Forensics face detection from gans [34] M. S. Rana and A. H. Sung, “Deepfakestack: A deep ensemble-
using convolutional neural network,” ISITC, vol. 2018, pp. 376–379, based learning technique for deepfake detection,” in 2020 7th IEEE
2018. international conference on cyber security and cloud computing
[14] X. Xuan, B. Peng, W. Wang, and J. Dong, “On the generalization of (CSCloud)/2020 6th IEEE international conference on edge computing
gan image forensics,” in Chinese conference on biometric recognition. and scalable cloud (EdgeCom). IEEE, 2020, pp. 70–75.
Springer, 2019, pp. 134–141. [35] H. H. Nguyen, F. Fang, J. Yamagishi, and I. Echizen, “Multi-task
[15] P. Yang, R. Ni, and Y. Zhao, “Recapture image forensics based on learning for detecting and segmenting manipulated facial images and
laplacian convolutional neural networks,” in International Workshop on videos,” in 2019 IEEE 10th International Conference on Biometrics
Digital Watermarking. Springer, 2016, pp. 119–128. Theory, Applications and Systems (BTAS). IEEE, 2019, pp. 1–8.
[16] B. Bayar and M. C. Stamm, “A deep learning approach to universal [36] Z. Guo, L. Hu, M. Xia, and G. Yang, “Blind detection of glow-based
image manipulation detection using a new convolutional layer,” in facial forgery,” Multimedia Tools and Applications, vol. 80, no. 5, pp.
Proceedings of the 4th ACM workshop on information hiding and 7687–7710, 2021.
multimedia security, 2016, pp. 5–10.
[17] T. Zhao, X. Xu, M. Xu, H. Ding, Y. Xiong, and W. Xia, “Learning self-
consistency for deepfake detection,” in Proceedings of the IEEE/CVF
international conference on computer vision, 2021, pp. 15 023–15 033.
[18] D. Afchar, V. Nozick, J. Yamagishi, and I. Echizen, “Mesonet: a compact
facial video forgery detection network,” in 2018 IEEE international
workshop on information forensics and security (WIFS). IEEE, 2018,
pp. 1–7.
[19] E. Sabir, J. Cheng, A. Jaiswal, W. AbdAlmageed, I. Masi, and P. Natara-
jan, “Recurrent convolutional strategies for face manipulation detection
in videos,” Interfaces (GUI), vol. 3, no. 1, pp. 80–87, 2019.
[20] Y. Li, M.-C. Chang, and S. Lyu, “In ictu oculi: Exposing ai created fake
videos by detecting eye blinking,” in 2018 IEEE International workshop
on information forensics and security (WIFS). IEEE, 2018, pp. 1–7.
[21] J. Donahue, L. Anne Hendricks, S. Guadarrama, M. Rohrbach, S. Venu-
gopalan, K. Saenko, and T. Darrell, “Long-term recurrent convolutional
networks for visual recognition and description,” in Proceedings of the
IEEE conference on computer vision and pattern recognition, 2015, pp.
2625–2634.