ABSTRACT
The field of AI-generated deepfake video detection is dynamic and evolving, with
researchers continuously exploring new methodologies, refining existing techniques, and
addressing challenges to create more accurate and reliable detection systems. Collaboration
between academia, industry, and policy makers is emphasised to navigate the ethical,
technical, and societal implications of these technologies.
The growing computation power has made the deep learning algorithms so powerful that
creating an indistinguishable human synthesised video popularly called as deep fakes has
become very simple. Possible situations where these convincing face-swapped deep fakes
might be employed include fabricating political events and false acts of terrorism,
generating revenge porn content, and engaging in blackmail schemes against individuals.
These scenarios highlight the potential misuse and harm associated with realistic deepfake
technology.In this project, we outline a novel deep learning approach designed to
successfully differentiate between artificially generated fake videos and authentic ones.
Our aim is to leverage Artificial Intelligence (AI) as a means to counteract the deceptive
capabilities of Artificial Intelligence (AI). Our system uses an XceptionNet combined with
LSTM to decide whether the video is deep fake or real. To emulate the real time scenarios
and make the models perform better on real time data, we evaluate our method on large
amount of balanced and mixed dataset prepared by mixing the various available dataset
like Face-Forensic [1], Deep Face detection challenge [2], Kaggle and Celeb-DF [3].
We also show how our system can achieve competitive results using a very simple
and robust approach.
Keywords: Artificial Intelligence(AI), Deep Learning (DL), Extreme Inception(Xception),
Long-Short Term Memory (LSTM), Machine Learning(ML),.
iii
Table of Contents
Title Page
APPROVAL LETTER i
ACKNOWLEDGEMENT ii
ABSTRACT iii
Table of Contents iv
List of Table v
List of Appendices vi
List of Abbreviations/Acronyms vii
CHAPTER 1 INTRODUCTION 1
1.1 Background 1
1.2 Motivation 2
1.3 Statement of the Problem 3
1.4 Project objective 3
1.5 Significance of the study 3
CHAPTER 2 LITERATURE REVIEW 4
CHAPTER 3 REQUIREMENT ANALYSIS 6
3.1 System Requirements 6
3.1.1 Functional Requirement 6
3.1.2 Non-Functional Requirement 7
3.1.3 Hardware Requirement 8
3.1.4 Software Requirement 8
CHAPTER 4 SYSTEM DESIGN AND ARCHITECTURE 10
4.1 System Architecture 10
4.2 Flowchart 11
iv
4.3 DFD Level 0 12
4.4 DFD Level 1 13
4.5 Use Case Diagram 14
4.6 ER Diagram 15
CHAPTER 5 METHODOLOGY 16
5.1 Software Development Model 16
5.2 XCEPTION 18
5.3 LSTM(Long short-term memory) 23
CHAPTER 6 RESULT AND ANALYSIS 25
CHAPTER 7 CONCLUSION,LIMITATION AND FUTURE
ENHANCEMENT 29
7.1 Conclusion 29
7.2 Limitation 29
7.3 Future Enhancement 29
REFERENCES 30
v
List of Figures
Title Page
Fig 4.1: System Architecture 10
Fig 4.2: Flowchart 11
Fig 4.3: DFD Level 0 12
Fig 4.4: DFD Level 1 13
Fig 4.5: Use Case Diagram 14
Fig 4.6: ER diagram 15
Fig 5.1: Agile Model 16
Fig 5.2: Xception 18
Fig 5.3: LSTM(Long short-term Memory) 23
Fig 6.1: Output(Real) 27
Fig 6.1: Output(Fake) 28
vi
List of Abbreviations/Acronyms
AI: Artificial Intelligence
CNN: Convolution Neural Network
DL: Deep Learning
ML: Machine Learning
Xception: Extreme Inception
vii
CHAPTER 1
INTRODUCTION
An "AI Generated Fake video Detector" web application is a digital tool powered by
artificial intelligence (AI) that aims to identify and flag videos that have been artificially
generated, manipulated, or edited using AI-based algorithms or video editing software.
The primary purpose of such an application is to combat the spread of misleading or
falsified visual content across digital platforms. It assists users, fact-checkers, journalists,
and online platforms in verifying the credibility of videos, aiding in the detection of
misinformation and enhancing the authenticity of visual content.
1.1 Background
AI-generated fake videos have the potential for misinformation and the erosion of trust in
visual content. Here are several specific drawbacks associated with AI-generated fake
videos. They are:
Misleading Information:AI-generated fake videos can mislead viewers by presenting false
or manipulated visual information. This misinformation can contribute to
misunderstandings, false narratives, and confusion, affecting decision-making and
perceptions.
Impact on Trust:The proliferation of fake videos can erode trust in the authenticity of visual
content online. Ascertaining whether a video is genuine or manipulated becomes
challenging, leading to skepticism and reduced trust in online sources.
Spreading False Narratives:Fake videos, especially when shared extensively on social
media or news platforms, can perpetuate false narratives. They can influence opinions,
public discourse, and even policy decisions based on inaccurate information.
1
Legal and Regulatory Challenges:Addressing the misuse of AI-generated fake videos poses
legal and regulatory challenges, as existing laws might not adequately cover these new
forms of digital manipulation.
1.2 Motivation
Latest rise of manipulated and fake videos across social media and digital platforms has
led to widespread misinformation. Detecting these videos helps in curbing the spread of
false narratives and misleading information.Promoting the responsible and ethical use of
AI technology is essential. Developing tools that detect AI-generated fake videos
encourages ethical considerations and responsible usage of advanced technologies.
Overall, the motivation behind creating an AI-generated fake video detector web
application is to address the challenges posed by the proliferation of manipulated visual
content and to promote a more informed, trustworthy, and safe online environment.
1.3 Statement of the Problem
In an era where Advanced AI technologies facilitate the creation and dissemination of
increasingly convincing fake videos, there exists a critical need for a robust and accessible
web application capable of swiftly and accurately detecting AI-generated or manipulated
visuals. The proliferation of misleading videos across digital platforms poses a substantial
threat to information credibility, public trust, and the ability to discern authentic content.
Addressing this challenge requires the development of an intelligent, user-friendly tool that
empowers individuals, journalists, fact-checkers, and online platforms to identify and
mitigate the impact of AI-generated fake videos, thereby fostering a more trustworthy and
informed digital environment.
So, From the above we feel the need of “AI Generated Fake video Detector” Web
Application.
1.4 Project objective
To develop a deepfake video detection web application.
2
1.5 Significance of the study
An AI-generated deepfake video detection web application holds significant importance
for various beneficiaries and can benefit them in several ways:
General Public: Provides the public with a tool to verify the authenticity of videos
encountered online, fostering a more informed and discerning audience. It helps individuals
avoid falling victim to misinformation, enhancing their digital literacy and critical thinking.
Online Platforms and Social Media: Helps online platforms implement measures to identify
and mitigate the spread of fake videos on their platforms, thereby fostering a more
trustworthy and responsible online community.
Fact-Checkers and Researchers: Supports fact-checking organisations and researchers by
streamlining the process of identifying manipulated visuals. This accelerates the
verification process and contributes to more accurate analyses of information.
Individuals Targeted by Manipulated videos: Protects individuals or groups depicted in
manipulated videos from potential harm, defamation, or false representation by swiftly
identifying and addressing the distribution of fake visuals.
By offering a reliable means of detecting AI-generated fake videos, this web application
contributes to a more trustworthy digital environment, empowers users to make informed
decisions, and aids various sectors in ensuring accuracy, credibility, and security in the use
of visual content online.
3
CHAPTER 2
LITERATURE REVIEW
In recent years there have been astonishing advances in AI-based synthetic media
generation.Thanks to deep learning methods it is now possible to generate visual data with
a high level of realism. This is especially true for human faces. Thus a lot of research has
been conducted on detecting these AI generated videos with the help of AI itself.
One of the major neural networks used for generating synthetic videos is GAN(Generative
Adversarial Network).The visual appearance of the videos generated by the latest GAN
architectures is so realistic that it deceives even the experienced and attentive observer.
This raises major concerns on the possible malicious use of such tools.Despite their high
visual quality, GAN videos are characterized by specific artifacts left from the generation
process that can be used to develop effective tools for their detection. In some cases, their
synthetic origin can be identified by visual inspection due to the presence of semantic
inconsistencies, such as color anomalies or lack of symmetries. More generally, these
videos present invisible artifacts, closely linked to the architecture of the generative
network, which can be extracted through appropriate processing steps. These artifacts
represent very strong clues, which can be exploited even when synthetic videos appear
perfectly realistic.[13]
D. Guera and J. Delp,[6] have used a recognition pipeline to automatically detect deep
fakes. They have proposed a two step analysis. During the first stage, features are extracted
at frame level using CNN. The second stage consists of RNN which will capture erratic
frames introduced due to the face swapping process. The dataset contains 600 videos
collected from various online sources. The accuracy achieved by their model is 94 percent.
Y. Li, MC. Chang and S. Lyu [9] have introduced a new system of exposing deep fakes
based on the eye blinking which are generated using neural networks. The paper focused
on analyzing the eye blinking in the video as it is a natural signal and it cannot be presented
well in the synthesized media. In the method the videos have been first preprocessed to
4
locate face area in each frame, then a Long Term Recurrent Convolutional
Network(LRCN) finds out temporal incongruity.
Adil Mohammed Parayil, Ameen Masood V, Muhammed Ajas P, Tharun R, and Usha K
[14] employ deep learning for deep fake detection, utilizing Convolutional Neural Network
(CNN) with the Xception network and ytLSTM for Recurrent Neural Network (RNN). The
algorithm recognizes spatial features and temporal inconsistencies between frames. Models
are trained and validated on three standard datasets, resulting in efficient deepfake
prediction with minimal computational time and nominal accuracy.
Franc¸ois Chollet [5] conducted the evaluation which employed a single Xception model
with a single crop of input images, comparing its performance to Inception V3 on ImageNet
and JFT datasets. Xception demonstrated marginal superiority over Inception V3 on
ImageNet, with a 4.3% relative improvement on JFT. The classification performance on
ImageNet revealed that Xception outperformed VGG-16, ResNet-152, and Inception V3,
achieving a Top-1 accuracy of 0.790 and Top-5 accuracy of 0.945. Notably, Xception
showcased a more substantial performance boost on JFT, potentially attributed to Inception
V3 being tailored for ImageNet, whereas Xception was not specifically tuned for JFT. Fine-
tuning Xception's hyperparameters for ImageNet could yield further enhancements.
So, From various literature surveys exploring diverse neural networks and deep learning
techniques, Convolutional Neural Networks (CNNs), specifically the Xception network
and Long Short-Term Memory (LSTM), emerge as standout tools. These models
demonstrate robustness by employing transfer learning to detect DeepFakes with
remarkable accuracy, facilitated by the ingestion of extensive training datasets.
5
CHAPTER 3
REQUIREMENT ANALYSIS
3.1 System Requirements
A requirement is a feature that the system must have or a constraint that must be accepted
by the client. Requirement Engineering aims at defining the requirements of the system
under construction. Requirement Engineering includes two main activities: requirement
elicitation which results in the specification of the system that the client understands and
analysis which results in an analysis model that the developer can unambiguously interpret.
A requirement is a statement about what the proposed system will do.
Requirements can be divided into two major categories:
3.1.1 Functional Requirement
Video Analysis:
Objective: Detect manipulated videos.
Details: Implement algorithms to analyze videos for signs of manipulation, including
inconsistencies in lighting, shadows, and textures.
Real-time Processing:
Objective: Enable real-time analysis.
Details: Minimize processing time to ensure quick identification of fake videos, suitable
for real-time applications like social media.
Integration:
Objective: Integrate with various platforms.
Details: Provide API integration for seamless collaboration with different applications and
services.
6
User Interaction:
Objective: User-friendly interface.
Details: Create an intuitive interface for users to interact with the system, view analysis
results, and provide feedback
Real-time processing with low latency.
Scalable architecture for handling a large volume of videos.
API integration for collaboration with various platforms.
User-friendly interface for interaction and feedback.
3.1.2 Non-functional Requirement
Accuracy:
Objective: High precision in detection.
Details: Minimise false positives and false negatives to ensure accurate identification of
manipulated videos.
Scalability:
Objective: Handle a large volume of videos.
Details: Design a scalable system architecture to accommodate a growing number of video
analysis requests.
Reliability:
Objective: Consistent performance.
Details: Ensure the system performs reliably across various types of manipulated videos
and under different conditions.
Security:
Objective: Protect user data.
7
Details: Implement security measures to safeguard user privacy during the video analysis
process.
Adaptability:
Objective: Adapt to new techniques.
Details: Design the system to adapt to emerging methods of video manipulation, staying
current with technological advancements.
Compliance:
Objective: Adherence to legal and ethical standards.
Details: Ensure the system complies with relevant laws and ethical standards, addressing
concerns related to user privacy and data protection.
Documentation:
Objective: Comprehensive documentation.
Details: Provide thorough documentation for administrators, developers, and end-users,
covering system functionality, configuration, and usage.
3.1.3 Hardware Requirements
• Processor: Intel Core i5 or higher
• RAM: 4 GB or higher
• Internet Access
3.1.4 Software Requirements
● Programming Language: Python
● Web Framework: Flask
● Machine Learning Libraries: TensorFlow, Keras
● Frontend Development: Flask,HTML
3.1.4.1 Programming Languages
• Python
3.1.4.2 Web Browser
• Any Chromium based Web Browser
3.1.4.3 Libraries
8
• Keras
• NumPy
• Pandas
• Flask
3.1.4.4 OS
• Windows,macOS
3.1.4.5 IDE
• VS Code
• Jupyter Notebook
• Google colab
9
CHAPTER 4
SYSTEM DESIGN AND ARCHITECTURE
4.1 System Architecture
Fig 4.1:System Architecture
10
4.2 Flowchart
Fig 4.2:Flowchart
11
4.3 DFD Level 0
Fig 4.3:DFD Level 0
12
4.4 DFD Level 1
Fig 4.4:DFD Level 1
13
4.5 Use Case Diagram
Fig 4.5 : Use Case Diagram
14
4.6 ER Diagram
Fig 4.6: ER Diagram
15
CHAPTER 5
METHODOLOGY
Section 5.1 Software development model
We have chosen the Agile model as our software development model.The Agile model is
an iterative and customer-centric approach to software development which involves short
development cycles, frequent customer collaboration, and a focus on delivering
incremental value.As any of the AI related works is of dynamic nature which requires
continuous refinement and real world-feedback,This model allows to respond quickly to
changing requirements, ensuring that the development process remains flexible and aligned
with the evolving needs of the project.
Fig 5.1: Agile Model
16
Step 1:Requirements Gathering
The requirements of the project are defined in this phase. The time and effort required for
the project should also be discussed. By analysing this information, we will be able
determine the system's economic and technical feasibility.
Step 2:Design the Re-equipments
In this phase,we analyse the system architecture and flow diagrams to further define the
requirements to show the work of new features and show how it will apply to the existing
system.
Step 3:Development/ iteration:
When the requirements are defined, the working on the project begins which aims to deploy
a working product. The product will undergo various stages of improvement, so it includes
simple, minimal functionality.
Step 4:Testing:
In this phase, The product's performance is examined and looked out for the bug and errors
Step 5:Deployment:
In this phase, A product for the user's work environment is issued.
Step6:Feedback:
After releasing the product, the last step is feedback. In this, feedback about the product is
received and worked through the feedback.
17
Section 5.2 Xception
Convolutional Neural Networks (CNNs) have revolutionised the field of computer vision
by delivering remarkable performances in various image-related tasks. One such
innovation in the architecture of CNNs is the XceptionNet, introduced by François Chollet
in 2017. XceptionNet combines the strengths of depthwise separable convolutions and
residual connections, resulting in a highly efficient and effective model. In this article, we
will dive deeper into the XceptionNet architecture and explore its unique characteristics
and advantages.
In Xception, The data first goes through the entry flow, then through the middle flow which
is repeated eight times, and finally through the exit flow. Note that all Convolution and
Separable Convolution layers are followed by batch normalization.
Fig 5.2 :Architecture of Xception
(Source: François Chollet. "Xception: Deep Learning with Depthwise Separable
Convolutions." )
18
Entry Flow: The Entry Flow serves as the initial component of XceptionNet and consists
of three stacked convolutional blocks. Each convolutional block contains a combination of
depthwise separable convolutions, residual connections, and max pooling. This component
extracts essential features from the input image and reduces its spatial dimension.
Middle Flow: The Middle Flow is responsible for processing the extracted features in a
highly efficient manner. It consists of eight convolutional blocks, each having a similar
structure. These blocks enhance the feature representation by incorporating residual
connections, depthwise separable convolutions, and skip connections. The skip
connections facilitate the direct flow of information across different blocks and foster the
propagation of gradients during training.
Exit Flow: The Exit Flow is the final component of XceptionNet and focuses on the
efficient classification of the input image. It comprises two convolutional blocks, followed
by global average pooling and a fully connected layer. The global average pooling reduces
the spatial dimensions, while the fully connected layer performs the actual classification
into various classes.
Depthwise separable convolutions consist of two separate operations – depthwise
convolutions and pointwise convolutions. These two operations are applied sequentially.
1) Depthwise convolution
Depthwise Convolution is a first step in which instead of applying convolution of size d ×
d × C , we apply a convolution of size d × d × 1 . In other words, we don’t make the
convolution computation over all the channels, but only 1 by 1. Here is an illustration of
the Depthwise convolution process :
19
Fig 5.3:Illustration of the Depthwise convolution process
(Source:https://maelfabien.github.io/assets/images/XCeption)
This creates a first volume that has size K × K × C , and not K × K × N as before. Indeed,
so far, we only made the convolution operation for 1 kernel /filter of the convolution, not
for N of them. This leads us to our second step.
2) Pointwise Convolution
Pointwise convolution operates a classical convolution, with size 1 × 1 × N over the K ×
K × C volume. This allows creating a volume of shape K × K × N. Here is an illustration
of the Pointwise Convolution :
Fig 5.4:Illustration of the Point Convolution
(Source:https://maelfabien.github.io/assets/images/XCeption2)
20
The number of operations is reduced by a factor proportional to 1/N.
Implementation
Xception offers an architecture that is made of Depthwise Separable Convolution blocks +
Max Pooling, all linked with shortcuts as in ResNet implementations.
The specificity of Xception is that the Depthwise Convolution is not followed by a
Pointwise Convolution, but the order is reversed, as in this example :
Fig 5.5: Depthwise Separable Convolution blocks + Maxpooling
(Source:https://maelfabien.github.io/assets/images/XCeption3)
21
ReLU
The rectified linear activation function, or ReLU for short, is a piecewise linear function
that, if the input is positive, outputs the input directly; else, it outputs zero. Because a model
that utilizes it is quicker to train and generally produces higher performance, it has become
the default activation function for many types of neural networks.At the end of CNN, there
is a Fully connected layer of neurons. As in conventional Neural Networks, neurons in a
fully connected layer have full connections to all activations in the previous layer and work
similarly. After training, the feature vector from the fully connected layer is used to classify
images into distinct categories. Every activation unit in the next layer is coupled to all of
the inputs from this layer. Overfitting occurs because all of the parameters are occupied in
the fully-connected layer. Overfitting can be reduced using a variety of strategies, including
dropout.Soft-Max is an activation layer that is typically applied to the network’s last layer,
which serves as a classifier. This layer is responsible for categorizing provided input into
distinct types. A network’s non-normalized output is mapped to a probability distribution
using the softmax function.
Section 5.3: LSTM(Long short-term memory)
LSTM networks are an extension of recurrent neural networks (RNNs) mainly introduced
to handle situations where RNNs fail.The basic difference between the architectures of
RNNs and LSTMs is that the hidden layer of LSTM is a gated unit or gated cell. It consists
of four layers that interact with one another in a way to produce the output of that cell along
with the cell state. These two things are then passed onto the next hidden layer. Unlike
RNNs which have got only a single neural net layer of tanh, LSTMs comprise three logistic
sigmoid gates and one tanh layer. Gates have been introduced in order to limit the
22
information that is passed through the cell. They determine which part of the information
will be needed by the next cell and which part is to be discarded. The output is usually in
the range of 0-1 where ‘0’ means ‘reject all’ and ‘1’ means ‘include all’.
Fig 5.5: LSTM architecture
(Source: https://miro.medium.com/v2/resize:fit:875/1*Mb_L_slY9rjMr8-IADHvwg)
Forget Gate
The forget gate in a recurrent neural network helps decide which information from the past
and present should be kept and which should be discarded. It does this by using weights
and biases to transform inputs, followed by a simple decision: keep (output 1) or discard
(output 0) the information.
Input gate
The addition of useful information to the cell state is done by the input gate. First, the
information is regulated using the sigmoid function and filter the values to be remembered
23
similar to the forget gate using inputs h_t-1 and x_t. Then, a vector is created using the tanh
function that gives an output from -1 to +1, which contains all the possible values from
h_t-1 and x_t. At last, the values of the vector and the regulated values are multipliedto
obtain useful information.
Output gate
The task of extracting useful information from the current cell state to be presented as
output is done by the output gate. First, a vector is generated by applying the tanh function
on the cell. Then, the information is regulated using the sigmoid function and filtered by
the values to be remembered using inputs h_t-1 and x_t. At last, the values of the vector
and the regulated values are multiplied to be sent as an output and input to the next cell.
24
CHAPTER 6
RESULT AND ANALYSIS
By harnessing the power of convolutional neural networks (CNNs) like Xception and
recurrent neural networks (RNNs) like LSTM, our web application aims to accurately
identify manipulated facial images and distinguish them from genuine ones.
Throughout this chapter, we will delve into the output produced by our detection system
through screenshots of ,
1. Model Total Loss vs Total Validation Loss
Fig: Model Total Loss vs Total Validation Loss
2. Model Total Accuracy vs Total Validation Loss
25
Fig: Model Total Accuracy vs Total Validation Loss
3. Model Accuracy
Fig: Model Accuracy
26
4. Testing Loss and Testing Accuracy
Fig: Testing Loss and Testing Accuracy
Output from the user side:
Fig: Output(Real)
27
Fig: Output(Fake)
28
CHAPTER 7
CONCLUSION, LIMITATIONS AND FUTURE ENHANCEMENT
Conclusion
Finally, We would like to conclude that the result of our “AI GENERATED VIDEO
DETECTOR” is better. Using XceptionNet and LSTM, we achieved overall accuracy of
86.25% , along with testing accuracy of 83.87%. And, using our system for deepfake
detection system is one effective approach. Also, our system is useful in the real world.
Limitation
Some of the limitations that we see are listed below as:
1. Generalization to New Deepfake Variants are bit challenging.
2. Incompatible with mobile devices.
FUTURE ENHANCEMENT
1. Using the latest release of ResNet 152 could be the best for deepfake video
detection.
2. Responsive GUI for all the devices.
3. Continuous monitoring and updating.
29
REFERENCES
[1] https://cs230.stanford.edu/projects_spring_2020/reports/38857501
[2] Deepfake detection challenge dataset: https://www.kaggle.com/c/deepfake-detection
challenge/data Accessed on 26 March 2020
[3] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi and Siwei Lyu “Celeb-DF: A Large-scale
Challenging Dataset for DeepFake Forensics” in arXiv:1909.12962
[4] Nagothu D, Xu R, Chen Y, Blasch E, Aved A. Deterring Deepfake Attacks with an Electrical
Network Frequency Fingerprints Approach. Future Internet. 2022; 14(5):125.
https://doi.org/10.3390/fi14050125
[5] Chollet, François. "Xception: Deep Learning with Depthwise Separable Convolutions."
arXiv preprint arXiv:1610.02357 (2016).
[6]D. Güera and E. J. Delp, "Deepfake Video Detection Using Recurrent Neural
Networks," 2018 15th IEEE International Conference on Advanced Video and Signal
Based Surveillance (AVSS), Auckland, New Zealand, 2018, pp. 1-6, doi:
10.1109/AVSS.2018.8639163.
[7] Aarti Karandikar.et al., International Journal of Advanced Trends in Computer Science
and Engineering, 9(2), March - April 2020, 1311 – 1315
[8] Zhiming Xie et al 2019 J. Phys.: Conf. Ser. 1395 012006
[9] Yuezun Li, Ming-Ching Chang and Siwei Lyu “Exposing AI Created Fake Videos by
Detecting Eye Blinking” in arXiv:1806.02877v2.
30
[10] Huy H. Nguyen, Junichi Yamagishi, and Isao Echizen “Using capsule networks to
detect forged videos and videos” in arXiv:1810.11215
[11] D. Güera and E. J. Delp, "Deepfake Video Detection Using Recurrent Neural
Networks," 2018 15th IEEE International Conference on Advanced Video and Signal
Based Surveillance (AVSS), Auckland, New Zealand, 2018, pp. 1-6.
[12] Umur Aybars Ciftci, ˙Ilke Demir, Lijun Yin “Detection of Synthetic Portrait Videos
using Biological Signals” in arXiv:1901.02212v2
[13] Gragnaniello, D., Marra, F., Verdoliva, L. (2022). Detection of AI-Generated
Synthetic Faces. In: Rathgeb, C., Tolosana, R., Vera-Rodriguez, R., Busch, C. (eds)
Handbook of Digital Face Manipulation and Detection. Advances in Computer Vision and
Pattern Recognition. Springer, Cham.
[14] Adil Mohammed Parayil, Ameen Masood V, Muhammed Ajas P, Tharun R, Usha K.
"Deepfake Detection Using Xception and LSTM." International Research Journal of
Modernization in Engineering Technology and Science, Volume 05, Issue 05, May 2023,
DOI: https://www.doi.org/10.56726/IRJMETS41123
31
