A Comprehensive Review on Deep Fake Detection

in Videos
Rajath R Pai, Rakeen Harmain Khan, Shreya N S, Suchith A
Dept. of Artificial Intelligence and Machine Learning
PES Institute of Technology and Management, Shivamogga, India
{rajathpai21, khan.harmain12, shreyagowda021, suchithstany324}@gmail.com

Abstract—The last few decades have seen a significant rise gaze, mouth position, open or closed eye color, injury, and
in Artificial Intelligence (AI) and Machine Learning (ML), indication of drug use. Manipulations of this kind give way
promoting the development of deepfake technology. Deepfakes to extremely realistic and believable changes in the original
are synthetic media created using AI techniques, altering audio,
images, and videos to appear authentic but are fabricated. media. A lot of techniques at an advanced level are applied
Em playing concepts like Generative Adversarial Networks for detection purposes. DenseNet, VGGNet, and ResNet are
(GANs), deepfake creation involves a competitive process where deep architectures that have proved very useful in image
one model produces forgeries while another aims to identify recognition and classification tasks, proving to be quite effec-
them. The consequences of deepfakes are extensive, ranging tive in the identification of subtle inconsistencies prevailing
from misinformation campaigns by terrorist organizations to
individual harm through fabricated pornography and financial in manipulated faces. The variation autoencoder learns the
scams. Detecting deepfakes presents a significant challenge, underlying distribution of the original data. In this way, it helps
prompting exploration into various methodologies. Convolutional in distinguishing between genuine and altered content. Capsule
Neural Networks (CNNs) stand as a prominent tool, with studies networks propose a new approach: they are particularly good
comparing different architectures such as DenseNet, VGGNet, at capturing spatial hierarchies between features, improving
ResNet, and custom CNNs. Additionally, autoencoders, partic
ularly Variational Autoencoders, offer utility in reducing data tampered region detection. One form of recurrent neural
dimensionality and editing facial features in images. Trans network work is LSTMs, which have been really good at tasks
fer learning, where pre-trained models like InceptionV3 are involving sequence prediction, making them very suitable for
adapted for deepfake detection, and multi-model approaches looking through video frames over time to detect manipula-
combining CNNs with Long-Short Term Memory Networks tions. All these techniques provide at one’s disposal a powerful
(LSTMs) have shown promise. Furthermore, novel techniques
like CapsuleNets and ensemble methods, which combine multiple toolkit for detecting and mitigating deepfake technologies.
base models, are explored for their efficacy. CapsuleNets facilitate
spatial inconsistency detection, while ensembling aims to enhance A. Deepfake Creation Techniques
model robustness. Two-stream networks, integrating spatial and The last few decades have seen rapid growth in Artificial In
frequency streams, offer a solution for handling low-quality telligence and Machine Learning. One of the technologies de
videos. Despite there being various advancements in detection
models, deepfake detection faces numerous challenges. Adverse veloped in these years is deepfake [1]. Deep fakes refer to syn
serial attacks targeting detection models, the emergence of new thetic media that are created using artificial intelligence tech
deepfake variants evading existing detection mechanisms, and niques, to manipulate or fabricate audio, images, and videos
computational resource requirements hindering real-time imple to make them appear real but are fake [2], [3]. DeepFake
mentation are significant hurdles. Moreover, concerns regarding uses the concept of generative adversarial networks (GANs), in
privacy, fairness, and biases in training data persist. As deepfake
generation technology progresses, detection models may struggle which two deep learning models compete [38], [40], [45]. One
to keep pace, requiring ongoing research and adaptation. While model gets trained on real data and tries to create forgeries;
detection models play a crucial role, a comprehensive approach meanwhile, the other strives to detect the forgery. The forger
to addressing deepfake-related issues is imperative. This involves keeps on creating better and better fakes until the other model
not only refining detection techniques but also fostering public is unable to detect the forgery [4]. Deepfakes can act as a
awareness, implementing regulatory measures, and promoting
responsible usage of AI-generated content. By addressing these powerful weapon for terrorist organizations that may create
versatile challenges, society can reduce the negative impacts of fake content about government officials to misguide the public
deepfake technology. [5]. Deepfakes are also used to create fake pornography, cause
Index Terms—Machine learning, Deep fake, Convolutional individual harm by tarnishing their image, and financial fraud
Neural Network, Generative Adversarial Networks. [6].
I. I NTRODUCTION B. Detection Techniques
Deepfakes in the field of face manipulation can repre- Face manipulation in deep fake involves modifying facial
sent modifications such as changes in age, gender, ethnicity, attributes such as age, gender, ethnicity, morphing, attractive
attractiveness, skin color or texture, color, style, or length ness, skin color or texture, hair color, style or length, eyeglass,
of hair, spectacles, makeup, mustache, emotion, beard, pose, makeup, mustache, emotion, beard, pose, gaze, mouth open
or closed, eye color, injury and effects of drug use [7], [8]. F. Challenges, Ethical, and Social Implications
Some of the techniques used for the detection of manipulation Although there are many different approaches have been ex
in videos are DenseNet, VGGNet, ResNet,Variational Autoen plored to detect deepfakes, the technology to create the deep
coders,CapsuleNets and Long-Short Term Memory Networks fakes is still ahead of the detection models, this is due to ad
(LSTMs) which can be used to determine the manipulation versarial attacks that are done on the model by terrorist groups
in the video [9]. Some of the methods like CNN-based to decrease the accuracy of the detection model [24]. Other
approaches mainly focus on analyzing facial features frameby- than that, the existing deepfake detection models themselves
frame, while RCNN-based methods incorporate spatial and have some problems. They might not work well with new types
temporal information to improve detection accuracy. Detection of deepfakes or could be tricked by small and unnoticeable
of these manipulations can involve error level analysis (ELA) changes made by the creators [25]. Some models need a lot
and spatial rich models (SRM) to identify anomalies.[57] of computer power, making them hard to use in real-time or
in smaller groups [18]–[21]. They might be too com plicated
C. Optical Flow Fields for Deepfake Detection for most people to understand and use. There are worries
This [10] proposes optical flow fields to detect deep fake about privacy and fairness because these models can be used
manipulation in videos, optical flow is a vector that describes to spy on people or unfairly target them. Sometimes, they
the apparent relative motion between the observer (camera) might not work the same for everyone or might make mistakes
and the scene itself. In the initial stages, the frames of the because they were trained on biased data from a dataset [23].
video are cropped into 300 x 300 pixels that contain the face As deepfake generation technology gets better, these models
of the person using a dlib face detector which are then passed might struggle to keep up, and some people might use tricks
as input to a ResNet50(CNN) which determines if the frame to avoid getting caught. It’s also important to remember that
is tampered or original. just relying on these models might not solve all the problems
deepfakes cause. We need to think about other ways to deal
with these issues, like teaching people how to spot fake videos
D. CNN vs Autoencoders vs Variational Autoencoders
and making rules about when and how deepfakes can be used.
A comparative study of the use of convolutional neural II. RELATED WORK
networks (CNN) has been conducted on eight different CNN
models. Tree of these models were trained using the DenseNet A. Role of Deep Learning and GANs in Deepfake Creation
architecture (DenseNet121, DenseNet169, and DenseNet201); Deepfakes in videos and images are created using vari
two were trained using the VGGNet architecture (VGG16, ous technologies such as deep learning, machine learning,
VGG19); one was with the ResNet50 architecture, one with computer vision, and generative adversarial networks (GANs).
the VGGFace, and one with a bespoke CNN architecture [39], These technologies enable the manipulation and synthesis of
[58]. A custom model has also been implemented to determine media content to generate hyper-realistic fake videos and im
whether other models reflect their objectives. Amongst all the ages. Deep learning techniques, including convolutional neu-
models, VGGFace performed the best, with 99% accuracy. ral networks (CNNs) and long short-term memory (LSTM),
Besides, we obtained 97% from the ResNet50, 96% from are commonly used for deepfake detection and identification
the DenseNet201, 95% from the DenseNet169, 94% from the [24], [25]. GANs, specifically StyleGANv2, and StarGAN,
VGG19, 92% from the VGG16, 97% from the DenseNet121 are utilized for generating realistic deepfakes by learning the
model, and 90% from the custom model [11]. distribution of real face data and generating corresponding face
images [25], [26]. Additionally, advanced models like Multi
E. Deepfake Detection Datasets and Benchmarks Crise-Cross Attention and Inception-ResNet-v2 are employed
for detecting deepfakes, especially in the presence of face
There are several deepfake detection datasets and bench- masks [27]. These technologies play a crucial role in both
marks available, including Face Forensics, FaceForensics++ the creation and detection of deepfakes, highlighting the need
[57], CelebDF, Face2Face [21], [22], FaceSwap [22], [23], for continuous research and development in this field.
and CDDB, the Continual Deepfake Detection Benchmark
[23]. Each of these is a useful resource in trying to build B. Advanced Models for Deepfake Generation
robust detection techniques. The datasets have been derived Deepfake detection is increasingly important in the present
from a number of internet videos and images. FaceForensics++ digital era in which false information spreads fast and brings
contains high-resolution videos with easily detectable artifacts, many challenges to doubts about the credibility of online
while CelebDF contains very diverse videos of celebrities. content [28]. To solve this problem, researchers have come up
CDDB evaluates the models on their adaptability against with different approaches that are meant to detect falsehoods.
new deepfake techniques. Another popular option is that of Additionally, some of these techniques use sophisticated con
GANbased datasets, very realistic face generation and yet do volutional neural networks (CNNs) which are trained using
not exist. Such datasets make sure that the detection models get various datasets to recognize bogus visuals from genuine
trained on very diverse and challenging examples to enhance video presentations. Moreover, several other methods examine
their real-world effectiveness. movements within videos to identify an abnormality that could
indicate refinement [29]. Other visualizations utilize 3D CNNs E. Addressing Computational Challenges in Model Integra-
for better spatial-temporal feature analysis in videos leading tion
to improved recognition. Meanwhile, combining CNNs with Collaboration among researchers, policymakers, and ethi
Re current Neural Networks (RNNs) has helped improve the cists is vital to expand suggestions and rules that safeguard in
accuracy of some images detected on media files with slight opposition to the misuse of deepfake detection technologies
changes even introduced into them while ViViT architectures and mitigate potential societal harms. At the same time as
are excellent at discovering eye cheats and D4 methods can sizable progress has been made inside the area of deepfake de
identify even more subtle fake videos in general, there are tection, ongoing studies, and collaboration are essential to cope
many approaches [30], [47]. with remaining challenges and enhance the ultra-modern on
this important place. By always refining detection strategies,
C. Importance and Advancements in Deepfake Detection in addressing obstacles, and prioritizing moral issues, allowing us
this digital era to create a more secure and greater sincere digital environment.
Efforts are being made to refine these detection systems to Through concerted efforts, staying ahead of emerging threats
overcome challenges such as change detection in poor quality and make certain the integrity and authenticity of virtual con
or highly compressed videos and detection of subtle changes in tent material for the benefit of society as an entire.
images [31]. Developing test methods necessitates addressing F. Ethical Considerations and International Cooperation
broad and potentially ethical concerns such as confidentiality,
disclosure, and the importance of individual autonomy [32]. It’s vital to understand the broader societal implications of
Collaboration between security researchers and policy-makers deepfake detection beyond technical advancements. As these
is crucial for crafting guidelines and regulations to tackle these technologies emerge as more state-of-the-art, the ability for
issues. Despite notable advancements in deepfake detection, misuse and harm will increase [53], [57]. Therefore, proac
numerous challenges persist [33]. tive measures should be taken to mitigate risks and protect
people’s rights and privacy [46]. This includes imposing ro
D. Overcoming Limitations in Detection Techniques bust privacy protections and transparency measures in deep
fake detection structures to ensure that private data is han
Optimizing research methods, addressing limitations, and dled responsibly and ethically. Public recognition and educa
prioritizing ethical considerations are important steps in creat tional tasks are essential to empower individuals to understand
ing a safe and trustworthy digital environment through ongoing and mitigate the effect of deepfake media. Promoting media
research and collaboration, allowing us to stay ahead of literacy and critical thinking skills allows people to better
emerging threats and ensure the authenticity and integrity of distinguish between genuine and manipulated content, thereby
digital content for years to come. As new techniques continue reducing the spread of misinformation and its harmful effects
to emerge, the imperative remains for constant adaptation on society arising from the introduction of the sword. Tack
and enhancement of detection methods to effectively mitigate ling multifaceted challenges requires a holistic approach that
evolving threats. Ongoing efforts are being made to refine and incorporates technical innovation, ethical considerations, and
optimize detection systems, addressing the unique challenges social awareness. By fostering collaboration across disci plines
they face [34]. The main obstacle is detecting manipula tions and stakeholders, we can develop comprehensive so lutions
in highly compressed low-quality videos or images [35]. Many that protect the integrity of digital content while re specting the
fake videos appear grainy or blurry to delib erately hide rights and freedoms of individuals. Through ongoing research,
imperfections, making it difficult for current images to be seen. education, and policy development, we can mitigate the risks
Apart from that, subtle changes in im ages, such as changes associated with deepfake media and build a more resilient and
occurring in foreground or background image recognition pose trustworthy digital ecosystem for future generations.
significant challenges to existing im ages Further research
and development are needed to develop it, and extensive G. Legal Frameworks and Regulations
testing of these recognition systems in real world situations The emergence of deepfake technology underscores the
is required to demonstrate reliability and effectiveness under need for robust legal frameworks and regulations to ad dress its
various conditions [36]. Researchers are working to develop potential misuse[49]. Legal measures can play a cru cial role
robust testing methods that demonstrate the difficulty of de- in deterring malicious actors from creating and dis seminating
tecting fake news under different light ing conditions, camera harmful deepfake content, while also providing av enues for
angles, and different media platforms where fake news can recourse for those impacted by its dissemination. By enacting
be shared if the accuracy and robustness of the system are laws that specifically target the creation and distribu tion
increased. Addressing the ethical impli cations of in-depth of deepfakes, policymakers can send a clear message that
identity searches is paramount to ensure that this technology such activities will not be tolerated and will be met with
is used responsibly and ethically [37]. It is important to strike ap propriate consequences [52]. International cooperation and
a balance between detecting and preventing the spread of collaboration are essential to effectively combine the global
fake news while upholding individuals’ rights to privacy and proliferation of deepfake media. Coordinated international
freedom of expression [42]. efforts are crucial due to the borderless nature of the internet
and the seamless sharing of digital content across different J. Neural Networks
jurisdictions. This may involve sharing best practices, intel
ligence, and resources to enhance detection capabilities and • Strengths: Capable of learning complex relationships
facilitate the swift removal of harmful deepfake content from in data, allowing them to capture small inconsistencies in
online platforms. Existing deepfake detection technologies deepfakes that might be missed by simpler models. Flexible
face several limitations that hinder their effectiveness in real architecture allows for various configurations, including con-
world situations. volutional neural networks (CNNs) for efficient image/video
analysis and recurrent neural networks (RNNs) for handling
temporal sequences in videos. Can be directly trained for
H. Limitations:
classification tasks, enabling them to output a probability of a
These limitations can be broadly categorized into three video being real or fake.
areas: generalizability, robustness, and detection limitations. • Weaknesses: Prone to overfitting, especially when deal-
1) Overfitting: Detectors often become overly focused on ing with limited training data, potentially leading to poor
specific patterns seen in their training data, leading to difficul- performance on unseen deepfakes. Require larger datasets
ties in identifying deepfakes created using different techniques and computational resources compared to autoencoders for
or with variations not present in the training set. effective training.
2) Limited dataset diversity: The training data used for These points have shown that deepfake detection has come
deepfake detection may not adequately represent the full far using various deep learning models, and that the most pre-
range of deepfakes encountered in the real world, leading to cise ones are based on CNN techniques, so it is really crucial in
challenges in generalizing to unseen scenarios. our battle against deepfakes. This project aims to build a smart
3) Novel techniques: Deepfake creators are constantly de- system using advanced technologies to detect and prevent
vel oping new methods, and existing detectors may struggle the spread of deceptive content, known as deepfakes, across
to identify forgeries created using these novel techniques that various types of data. By combining different techniques and
were not present in the training data. algorithms, the system will significantly improve its ability to
4) Degraded videos: The accuracy of deepfake detection accurately identify and counter synthetic media manipulation,
often decreases for low-quality or compressed videos, which ensuring a more trustworthy digital environment.
are commonly shared on social media platforms.
5) Detection limitations: Limited information extraction: K. Analysis and Discussion
Current methods may not effectively capture all the crucial
information from deepfakes, hindering their ability to accu- While discussing the use of different forms of deepfake
rately detect them. detection models, some things should be taken into account.
Firstly, the degree of complexity present in deepfakes is
6) Single-domain reliance: Most methods rely solely on
significant. Simpler cases can detect using autoencoders, while
visual information for detection, neglecting potential in sights
more complicated ones may need advanced feature learning
that could be gained from analyzing other domains like audio
abilities provided by neural networks. Secondly, training data
or text alongside the visuals.
availability matters a lot. For limited data situations such as
These limitations highlight the need for continuous research
these, autoencoders make a good option due to their less
and development in deepfake detection technologies to im
computational requirements. Also, the kind of work it takes to
prove their generalizability, robustness, and overall effective
identify something determines which model one could apply.
ness in the face of evolving deepfake creation techniques.
Autoencoders might be suitable for anomaly detection goals
Major technologies for deepfake detection are autoencoders
while direct classification of deepfakes may be better solved
and neural networks, each offering distinct advantages and
with neural networks. Additionally, costs associated with
disadvantages.
merging these models are not negligible. Lastly, a possible
future direction would involve combining the strengths of both
I. Autoencoders autoencoders and NNs. By this way, efficient feature learning
Strengths: Efficiently learn latent representations of genuine capabilities from autoencoders can coexist with robust clas
data, allowing them to identify inconsistencies in deepfakes sification ability from neural network to enhance or improve
that deviate from these learned patterns. Use ful for anomaly deep fake detection methodologies.
detection, as they can flag deviations from the expected recon-
struction of real data during the encoding-decoding process. III. ACKNOWLEDGMENT
Weaknesses: Efficiently learn latent representations of gen-
uine data, allowing them to identify inconsistencies in deep- We would like to express our gratitude to all of the
fakes that deviate from these learned patterns. Use ful for researchers, developers, lecturers and volunteers that have
anomaly detection, as they can flag deviations from the contributed to the fields of Deep Fake identification. Your ded-
expected reconstruction of real data during the encoding- ication and inventiveness has made a significant contribution
decoding process. to our knowledge and advancement in this area.
 R EFERENCES [23] Li, C., Huang, Z., Paudel, D. P., Wang, Y., Shahbazi, M., Hong, X.,
& Van Gool, L. (2023). A continual deepfake detection benchmark:
Dataset, methods, and essentials. In Proceedings of the IEEE/CVF
[1] Rana, M. S., Nobi, M. N., Murali, B., & Sung, A. H. (2022). Deepfake Winter Conference on Applications of Computer Vision (pp. 1339-
detection: A systematic literature review. IEEE Access, 10, 25494-25513. 1349).
[2] Wubet, W. M. (2020). The deepfake challenges and deepfake video [24] Neekhara, P., Dolhansky, B., Bitton, J., & Ferrer, C. C. (2021).
detection. Int. J. Innov. Technol. Explor. Eng, 9. Adversarial threats to deepfake detection: A practical perspective. In
[3] Chadha, A., Kumar, V., Kashyap, S., & Gupta, M. (2021). Deepfake: Proceedings of the IEEE/CVF Conference on Computer Vision and
an overview. In Proceedings of Second International Conference on Pattern Recognition (pp. 923-932).
Computing, Communications, and Cyber-Security: IC4S 2020 (pp. 557- [25] Gowrisankar, B., & Thing, V. L. (2024). An adversarial attack approach
566). Springer Singapore. for eXplainable AI evaluation on deepfake detection models. Computers
[4] Singh, A., Saimbhi, A. S., Singh, N., & Mittal, M. (2020). DeepFake Security, 139, 103684.
video detection: a time-distributed approach. SN Comput. Sci, 1, 212. [26] Xiao, S., Lan, G., Yang, J., Lu, W., Meng, Q., & Gao, X. (2023).
[5] Katarya, R., & Lal, A. (2020, October). A study on combating emerging MCSGAN: A Different Understanding for Generalization of Deep
threat of deepfake weaponization. In 2020 Fourth International Confer- Forgery Detection. IEEE Trans. Multimedia.
ence on I-SMAC (IoT in Social, Mobile, Analytics and Cloud)(I-SMAC) [27] Campbell, C., Plangger, K., Sands, S., Kietzmann, J., & Bates, K. (2022).
(pp. 485-490). IEEE. How deepfakes and artificial intelligence could reshape the advertising
[6] Tolosana, R., Vera-Rodriguez, R., Fierrez, J., Morales, A., & Ortega- industry: The coming reality of AI fakes and their potential impact on
Garcia, J. (2020). Deepfakes and beyond: A survey of face manipulation consumer behavior. J. Advert. Res., 62(3), 241-251.
and fake detection. Information Fusion, 64, 131-148. [28] Liu, J., Zhu, K., Lu, W., Luo, X., & Zhao, X. (2021). A lightweight 3D
[7] Akhtar, Z. (2023). Deepfakes Generation and Detection: A Short Survey. convolutional neural network for deepfake detection. Int. J. Intell. Syst.,
Journal of Imaging, 9(1), 18. 36(5), 2313-2329.
[8] Mirsky, Y., & Lee, W. (2021). The creation and detection of deepfakes: [29] Prashanth, K., & Kundu, H. (2021). An adaptive neural network ap-
A survey. ACM Comput. Surv., 54(1), 1-41. proach for deepfake detection. J. Ambient Intell. Humaniz. Comput., 12,
[9] Kharbat, F. F., Elamsy, T., Mahmoud, A., & Abdullah, R. (2019, 5437-5445.
November). Image feature detectors for deepfake video detection. In [30] Bondi, L., Cannas, E. D., Bestagini, P., & Tubaro, S. (2020, December).
2019 IEEE/ACS 16th International Conference on Computer Systems Training strategies and data augmentations in CNN-based deepfake
and Applications (AICCSA) (pp. 1-4). IEEE. video detection. In 2020 IEEE International Workshop on Information
[10] Caldelli, R., Galteri, L., Amerini, I., & Del Bimbo, A. (2021). Optical Forensics and Security (WIFS) (pp. 1-6). IEEE.
Flow based CNN for detection of unlearnt deepfake manipulations. [31] Suratkar, S., Kazi, F. (2023). Deep fake video detection using transfer
Pattern Recognition Letters, 146, 31-37. learning approach. Arabian J. Sci. Eng., 48(8), 9727-9737.
[11] Shad, H. S., Rizvee, M. M., Roza, N. T., Hoq, S. M., Monirujjaman [32] Ramadhani, K. N., Munir, R., & Utama, N. P. (2024). Improving
Khan, M., & Singh, Bourouis S. (2021). Comparative analysis of Video Vision Transformer for Deepfake Video Detection using Facial
deepfake image detection method using convolutional neural network. Landmark, Depthwise Separable Convolution and Self Attention. IEEE
Comput. Intell. Neurosci., 2021. Access.
[12] Makhzani, A., Shlens, J., Jaitly, N., Goodfellow, I., & Frey, B. (2015). [33] Yu, Z., Cai, R., Li, Z., Yang, W., Shi, J., & Kot, A. C. (2024).
Adversarial autoencoders. arXiv preprint arXiv:1511.05644. Benchmarking joint face spoofing and forgery detection with visual and
[13] Li, D., Zhang, M., Chen, W., & Feng, G. (2018, August). Facial attribute physiological cues. IEEE Trans. Dependable Secure Comput..
editing by latent space adversarial variational autoencoders. In 2018 [34] Leyva, R., Sanchez, V., Epiphaniou, G., & Maple, C. (2024). Detect-
24th International Conference on Pattern Recognition (ICPR) (pp. 1337- ing Face Synthesis Using a Concealed Fusion Model. arXiv preprint
1342). IEEE. arXiv:2401.04257.
[35] Nawaz, M., Javed, A., & Irtaza, A. (2023). ResNet-Swish-Dense54: a
[14] Suratkar, S., Johnson, E., Variyambat, K., Panchal, M., & Kazi, F. (2020,
deep learning approach for deepfakes detection. The Visual Computer,
July). Employing transfer-learning based CNN architectures to enhance
39(12), 6323-6344.
the generalizability of deepfake detection. In 2020 11th International
[36] Killi, C. B. R., Balakrishnan, N., & Rao, C. S. (2023). Deep Fake
Conference on Computing, Communication and Networking Technolo-
Image Classification Using VGG-19 Model. Ingénierie des Systèmes
gies (ICCCNT) (pp. 1-9). IEEE.
d’Information, 28(2).
[15] Sanghvi, B., Shelar, H., Pandey, M., & Sisodia, J. (2021, April). De-
[37] Waqas, N., Safie, S. I., Kadir, K. A., Khan, S., & Khel, M. H. K. (2022).
tection of machine generated multimedia elements using deep learning.
DEEPFAKE image synthesis for data augmentation. IEEE Access, 10,
In 2021 5th International Conference on Computing Methodologies and
80847-80857.
Communication (ICCMC) (pp. 1238-1243). IEEE.
[38] Zhang, T., Deng, L., Zhang, L., & Dang, X. (2020, August). Deep
[16] Nasar, B. F., Sajini, T., & Lason, E. R. (2020, December). Deepfake learning in face synthesis: A survey on deepfakes. In 2020 IEEE 3rd
detection in media files-audios, images and videos. In 2020 IEEE Recent International Conference on Computer and Communication Engineering
Advances in Intelligent Computational Systems (RAICS) (pp. 74-79). Technology (CCET) (pp. 67-70). IEEE.
IEEE. [39] Guarnera, L., Giudice, O., Guarnera, F., Ortis, A., Puglisi, G., Paratore,
[17] Guera, D., & Delp, E. J. (2018, November). Deepfake video detection A., ... & Battiato, S. (2022). The face deepfake detection challenge.
using recurrent neural networks. In 2018 15th IEEE International Con- Journal of Imaging, 8(10), 263.
ference on Advanced Video and Signal Based Surveillance (AVSS) (pp. [40] Fischer, E. M. Formulation of the Variational Autoencoder and Evidence
1-6). IEEE. Lower Bound and an Application to the MNIST Dataset.
[18] Sabir, E., Cheng, J., Jaiswal, A., AbdAlmageed, W., Masi, I., & Natara- [41] Zhao, Y., & Linderman, S. (2023, July). Revisiting structured variational
jan, P. (2019). Recurrent convolutional strategies for face manipulation autoencoders. In International Conference on Machine Learning (pp.
detection in videos. Interfaces (GUI), 3(1), 80-87. 42046-42057). PMLR.
[19] Mehra, A. (2020). Deepfake detection using capsule networks with long [42] Remya Revi, K., Vidya, K. R., & Wilscy, M. (2021). Detection of
short-term memory networks (Master’s thesis, University of Twente). Deepfake Images Created Using Generative Adversarial Networks: A
[20] Bonettini, N., Cannas, E. D., Mandelli, S., Bondi, L., Bestagini, P., & Review. In Second International Conference on Networks and Advances
Tubaro, S. (2021, January). Video face manipulation detection through in Computational Technologies: NetACT 19 (pp. 25-35). Springer Inter-
ensemble of CNNs. In 2020 25th International Conference on Pattern national Publishing.
Recognition (ICPR) (pp. 5012-5019). IEEE. [43] Ganguly, S., Mohiuddin, S., Malakar, S., Cuevas, E., & Sarkar, R.
[21] Kohli, A., & Gupta, A. (2021). Detecting deepfake, faceswap and (2022). Visual attention-based deepfake video forgery detection. Pattern
face2face facial forgeries using frequency CNN. Multimedia Tools Appl., Anal. Appl., 25(4), 981-992.
80, 18461-18478. [44] Hooda, A., Mangaokar, N., Feng, R., Fawaz, K., Jha, S., & Prakash,
[22] Liang, Y., Wang, M., Jin, Y., Pan, S., & Liu, Y. (2023). Hierarchical A. (2024). D4: Detection of Adversarial Diffusion Deepfakes Using
Supervisions with Two-stream Network for Deepfake Detection. Pattern Disjoint Ensembles. In Proceedings of the IEEE/CVF Winter Conference
Recognit. Lett.. on Applications of Computer Vision (pp. 3812-3822).
[45] Ganiyusufoglu, I., Ngo, L. M., Savov, N., Karaoglu, S., & Gevers, T.
(2020). Spatio-temporal features for generalized detection of deepfake
videos. arXiv preprint arXiv:2010.11844.
[46] Walia, S., Kumar, K., & Kumar, M. (2023). Unveiling digital image
forgeries using Markov based quaternions in frequency domain and
fusion of machine learning algorithms. Multimedia Tools Appl., 82(3),
4517-4532.
[47] Guera, D., & Delp, E. J. (2018, November). Deepfake video detection
using recurrent neural networks. In 2018 15th IEEE International Con-
ference on Advanced Video and Signal Based Surveillance (AVSS) (pp.
1-6). IEEE.
[48] Han, X. Y., & Lewis, A. S. (2020). Disk matrices and the proximal
mapping for the numerical radius. arXiv preprint arXiv:2004.14542.
[49] Coccomini, D., Caldelli, R., Falchi, F., & Gennaro, C. (2023). On the
Generalization of Deep Learning Models in Video Deepfake Detection.
Journal of Imaging, undefined.
[50] Wang, Y., Wang, J., Liang, Y., Chen, X., & Guo, Y. (2023). Deepfake
Video Detection via Facial Action Dependencies Estimation.
[51] Prasad, S. S., et al. (2022). Human vs. Automatic Detection of Deepfake
Videos Over Noisy Channels. In 2022 IEEE International Conference
on Multimedia and Expo (ICME). IEEE.
[52] Lin, Y., et al. (2020). Exploring generalization capability for video
forgery and detection based on generative adversarial network. In 2020
International Conference on Computational Science and Computational
Intelligence (CSCI). IEEE.
[53] Yang, J., Li, A., Xiao, S., Lu, W., & Gao, X. (2021). MTD-Net:
Learning to detect deepfakes images by multi-scale texture difference.
IEEE Trans. Inf. Forensics Security, 16, 4234-4245.
[54] Xu, Y., Zhang, D. D., & Sun, C. (2023, June). Frequency domain
deepfake detection based on two-stream neural network. In Fourteenth
International Conference on Graphics and Image Processing (ICGIP
2022) (Vol. 12705, pp. 910-918). SPIE.
[55] BR, S. R., Pareek, P. K., Bharathi, S., & Geetha, G. (2023, February).
Deepfake Video Detection System Using Deep Neural Networks. In
2023 IEEE International Conference on Integrated Circuits and Com-
munication Systems (ICICACS) (pp. 1-6). IEEE.
[56] Hu, Y., Zhang, X., & Ren, J. (2023). Automatic detection of deepfake
videos using convolutional neural networks in frequency domain. Mul-
timedia Tools Appl., 82(2), 1827-1847.
[57] Heidari, A., Jafari Navimipour, N., Dag, H., Unal, M. (2024). Deepfake
detection using deep learning methods: A systematic and comprehensive
review. Wiley Interdisciplinary Reviews: Data Mining and Knowledge
Discovery, 14(2), e1520.