Deepfake Detection Using CNN And DCGANS To

Drop-Out Fake Multimedia Content: A Hybrid

Approach
Kartik Bansal Shubhi Agarwal Narayan Vyas
Department of CSE Department of CSE Department of CSE
Chandigarh University Chandigarh University Chandigarh University
Mohali, Punjab, India Mohali, Punjab, India Mohali, Punjab, India
[email protected] [email protected] [email protected]
2023 International Conference on IoT, Communication and Automation Technology (ICICAT) | 979-8-3503-0282-0/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICICAT57735.2023.10263628

Abstract: The creation of DeepFakes, which are altered videos, Big fat companies like Facebook and Google face
audio, and photographs capable of disseminating false competition as their image and video resources must be
information and fake news and modifying sensitive records, is the validated and verified. These sites are the resources where
result of the rapid advancements in artificial intelligence and most people collect information, which could be misleading
machine learning. DeepFakes may also be used for interactive and unauthentic [6]. It needs to be significantly reduced,
learning and visual effects in entertainment and education. As a which can be done by applying a layering methodology that
result, numerous deep learning models, such as Convolutional helps to tackle regression and classification analysis [7].
Neural Networks (CNN) and Generative Adversarial Networks
(GAN), are being used for detection. DeepFakes detection and
removal have become essential challenges. Facebook AI's
Deepfake Detection Challenge (DFDC) dataset is invaluable for
developing and evaluating detection techniques. While it
represents serious risks, creating trustworthy detection techniques
might lessen their impact and enable investigation of their possible
beneficial applications. To ensure the authenticity and
dependability of multimedia information in the face of the ongoing
DeepFake threat, this paper emphasizes the importance of transfer
learning, deep learning, and optimization techniques in building
effective detection models. By doing this, we can stop the spread of
fake news and information, protect the public's trust, and promote
the moral and beneficial application of DeepFake technology
across various fields.

Keywords— Transfer Learning, Deep Learning,

Convolutional Neural Networks (CNN), Image Classification,
ImageNet, Optimization Techniques

I. INTRODUCTION
In this era of technology of leaving digital footprints, with
the rapid process in fields of Artificial Intelligence and deep
learning mechanisms, manipulation has taken a step forward
in the case of multimedia [1]. Creating fake images and videos
using image processing tools such as GNU Gimp and Adobe
Photoshop is a big problem. For example, they are the main
source of fake news and are often used to incite crowds.
Before acting on a false image, we must check its reality. The
original content of videos or images, specifically human faces,
are manipulated using GAN [2], which is unsupervised
learning where the model tries to create a replica of the
training set i.e. deform the image received, which can cause
malicious effects on one’s identity and can cause social
defamation [3]. DeepFake can be interpreted as synthesizing
the image from the landmarks available using Generator. Then
embedder tries to find the ground truth about the image. A
discriminator enters the picture to distinguish between these
images and predict the realism score [4].
The manipulation is not only happening for unlawful
actions, but DeepFake are largely used in entertainment, i.e.,
for animations, VFX used in movies wherein human faces
need to be defamed, etc. Also, DeepFakes find great uses in
education to make the learning procedure more efficient and
interactive [5].
Fig. 1. Process of DeepFake image generation

Authorized licensed use limited to: K K Wagh Inst of Engg Education and Research. Downloaded on August 21,2024 at 16:22:36 UTC from IEEE Xplore. Restrictions apply.
 Fig. 1 illustrates how GAN and CNN are used to create
DeepFake images. A generator network takes in a noise vector Worries About Deepfake
and outputs a fabricated picture. Both networks are then
backpropagated to reflect the new, more accurate evaluations Technology By Country
made by the discriminator network after it compares the false 40%
and actual images. This procedure is repeated until the
generated images are so lifelike that they can hardly be told 30%
apart from the real thing.
DeepFake are identified based on residual noise, warping, 20%
and blurring effects. When any image goes into auto-encoder
or GAN, the changes in the image reflect these criteria, which 10%
the developed model should work upon and differentiate
between the real and fake images [8]. Due to the presence of 0%
immense resources for verification purposes, it usually takes UK US Canada Global
thousands of GPU (General Processing Unit) hours to create a
Misinformation Political manipulation
new and modern computer vision right from the scrap [9].
Verification difficulty National security
II. LITERATURE REVIEW
In the research [10], DeepFake technology can potentially Fig. 2. Percentage of Related Worries about DeepFake Technology in
create realistic fake multimedia content that can be used for Different Countries
malicious purposes. As a result, detecting and removing This paper [15] surveys the research on DeepFake
DeepFake content has become a pressing issue. CNN and detection techniques for identifying fake photographs and
Deep Convolutional Generative Adversarial Networks videos of people's faces. The review compiles the current
(DCGANs) are two techniques used for DeepFake detection. DeepFake production methods and divides them into five
This work [11] uses optical flow features to provide a classes. Improvements to existing DeepFake datasets are also
hybrid CNN-LSTM model for video DeepFake detection. discussed in the study. The difficulties in creating and
The study explores the difficulties in detecting DeepFakes detecting DeepFakes are discussed, and the topic of creating
produced by deep generative algorithms like GAN using a generalized DeepFake detection model is analyzed.
conventional detection techniques. The suggested model This research [16] explores local motion and establishes
combines CNN and RNN architectures for classification and a dynamic inconsistency modelling framework to provide a
uses optical flow-based feature extraction to extract temporal novel method for detecting DeepFake movies. The suggested
data. The model exhibits encouraging results on open-source technique performs better on four benchmark datasets than
datasets like DFDC, FF++, and Celeb-DF with fewer sample the most recent competitors. The research identifies a flaw in
numbers. current methods that ignore local motions across consecutive
This study [12] suggests a deep learning strategy for frames and suggests a remedy to address this problem.
identifying DeepFake videos—hyper-realistic fake videos InceptionResnet v2 and Xception are used in the paper's
produced using sophisticated image editing methods. The [17] hybrid CNN model for DeepFake video detection to
model takes a layered approach, first using facial recognition extract frame-level characteristics. The model attained a high
networks to identify the subject, then CNN and LSTM layers f1-score, recall, and precision. The methodology uses
to extract facial features and look for face manipulation Kaggle's DFDC deep fake detection challenge for
between frames in temporal sequences. The research exposes experimental analysis.
the data compression-related limitations of conventional
picture forensics methods and makes a case for more TABLE I. ANALYSIS OF DIFFERENT ALGORITHMS
sophisticated methods to identify DeepFakes.
Study Method Remark Pros Limitation
This paper [13] surveys the current state of study on A
DeepFake detection and suggests an approach to fill the void combination
in the quest for a more universal solution. To dynamically of CNNs and
synthesize the most difficult forgeries to the present model, DCGANs
the suggested method involves synthesizing augmented CNN improved the Improved Computationally
[18] and detection detection expensive and
forgeries with a pool of forgery configurations and employing DCGAN accuracy of accuracy complex training
adversarial training. DeepFakes
compared to
This study [14] examines the background, methodology, using only a
and proposed approaches for detecting and evaluating threats CNN.
posed by DeepFake technology. The assessment emphasizes Proposed
the unanswered questions in the field and calls for improved method
DeepFake detection strategies. To detect DeepFakes, the achieved
DCGAN higher Higher
paper finds that CNNs are the most promising method. [19] and detection detection
Limited
scalability
CNN accuracy accuracy
compared to
using only a
CNN.
CNN Proposed High
Limited
[20] and method detection
generalizability
DCGAN achieved accuracy

Authorized licensed use limited to: K K Wagh Inst of Engg Education and Research. Downloaded on August 21,2024 at 16:22:36 UTC from IEEE Xplore. Restrictions apply.
 Study Method Remark Pros Limitation III. METHODOLOGY
high for
detection different A. Dataset Description
accuracy for types of
both face- DeepFakes To advance DeepFake detection research, Facebook AI
swapping launched the DFDC dataset in partnership with academic
and face partners [20]. The dataset contains over 100,000 videos
reenactment generated using several deepfake methods and techniques,
DeepFakes. including face-swapping, facial re-enactment, and lip-
MesoNet
outperformed
syncing.
traditional Both training and a validation dataset are included in the
Limited to
CNN models High
[21] MesoNet and achieved detection
detecting DFDC data set. There are 5,639 DeepFake films and 5,823
specific types of real videos in the training dataset, while there are 1,000
high accuracy
DeepFakes DeepFake videos and 1,000 real videos in the validation
accuracy in
detecting dataset. Actors and actresses of varied ages, genders, and
DeepFakes. ethnicities are represented in both datasets.
Capsule
Network The DFDC dataset additionally includes metadata, such as
showed the video's original source, DeepFake generation method, and
promising Requires more
results in Robust computational
type of edit. This data can be used to improve methods for
Capsule detecting against resources detecting deepfakes and learn more about them. This dataset
[22] is an important resource for academics and practitioners of
Network DeepFakes adversarial compared to
and was attacks traditional CNN deepfake identification due to its wide variety of challenging
robust models test cases, such as low-quality films and those with minor
against modifications. Research on the DFDC dataset has been so
adversarial
attacks. successful that new models and techniques for detecting
DeepFakes have been created.
Table I compares five studies published between 2021
and 2022 that propose various techniques for detecting TABLE II. DATASET DESCRIPTION
DeepFakes using deep learning architectures. The table lists Attribute
Type Size Description
the year of publication, study name, detection method, key Name
remark, advantage, and limitation for each study. Over DFDC dataset aids research on
Source Video 100,000 DeepFake detection using
The table indicates that using combinations of deep videos synthetic videos.
learning architectures can improve the accuracy of DeepFake DFDC dataset has diverse
Training 11,462
Video videos: 5,639 DeepFakes, 5,823
detection. The author [23] used a combination of CNNs and Dataset videos
real videos.
DCGANs to improve the detection accuracy of DeepFakes. Validation dataset has 1,000 real
Similarly, [24] used a combination of CNNs and DCGANs to Validation
Video
2,000 and DeepFake videos for testing,
achieve high accuracy in detecting different types of Dataset videos similar to training dataset but for
validation purposes.
DeepFakes. DFDC dataset videos created
using various DeepFake
In addition, the table shows that other deep learning DeepFake
Metadata
250
techniques such as facial
architectures, such as MesoNet and Capsule Networks, have Methods characters
reenactment, face-swapping, and
shown promising results in detecting DeepFakes [25]. The lip-syncing.
MesoNet outperformed traditional CNN models in detecting DFDC DeepFake videos are
50
Quality Metadata high quality, hard to detect
DeepFakes. characters
manually.
Fig. 3 and 4 below depict the trends of web and image DFDC dataset has a diverse
50 range of actors and actresses of
searches done on the most famous search engine i.e., Google Diversity Metadata
characters different ages, gender, and
and YouTube, for the past month while writing this review ethnicities.
paper [26]. It is shown how often the word “DeepFake” has 100
DFDC dataset used for
been searched on these platforms. Usage Research developing and evaluating
characters
DeepFake detection.
The DFDC dataset is summarized in Table II. Metadata
such as training and validation dataset sizes and information
about DeepFake techniques, quality, and diversity are
included. Research groups and organizations widely use the
DFDC dataset to create and test DeepFake detection
Fig. 3. Web Search techniques. This is difficult since DeepFake movies tend to be
of high quality, making them difficult for human observers to
spot.
B. Proposed Model
Deepfakes are fabricated news that can convince readers
Fig. 4. Image Search or viewers of an untruth. GAN has emerged as a potential
approach for detecting DeepFake. GANs can generate
synthetic DeepFake films that closely mimic genuine videos,
boosting the performance of DeepFake detection models on

Authorized licensed use limited to: K K Wagh Inst of Engg Education and Research. Downloaded on August 21,2024 at 16:22:36 UTC from IEEE Xplore. Restrictions apply.
 the DFDC dataset. Using input noise vectors, the GAN  Train the combined model on the extracted features
generator network creates new images, while the GAN  Evaluate the model on the testing set
discriminator network attempts to tell the fake ones from the  CNN combined loss function:
actual ones in the DFDC dataset. The generator network gets 𝐶𝑁𝑁𝑙𝑜𝑠𝑠 = ⁡ −𝑌 ∗ log(𝑝) − (1 − 𝑌) ∗ log(1 − 𝑝)
better at producing images that are aesthetically comparable
where 𝑋 is the input image, 𝑌 is the corresponding
to real photos as the two networks compete against each other
in an adversarial way. label (real or generated), and 𝑝 is the predicted
probability of the input image being real.
While GANs have demonstrated promising results in
enhancing the precision of DeepFake detection algorithms,
they still face obstacles that must be overcome. For instance,
GANs might create produced images with subtle visual
artefacts, making it hard to tell them from real photographs.
The high computational cost of GANs also hinders their
scalability in large-scale settings. Overall, GANs offer a
potent method for enhancing the precision of DeepFake
detection models, but more study is required to uncover their
full potential and overcome their limitations.
Furthermore, DeepFake detection depends not only on the
performance of the deep learning algorithms but also on the
quality of the input data. To improve the robustness and
generalizability of deep learning models, the training data
utilised for DeepFake detection must be diverse,
representative, and unbiased.
1) Pseduo Code
Step 1: Preprocessing
 Load the DeepFake dataset
 Split the dataset into training and testing sets
 Resize all images to the same size
 Normalize the pixel values of the images
Step 2: Training the DCGAN
 Define the generator and discriminator models
 Train the DCGAN on the real and fake images
 Save the trained generator model
 Training the DCGAN:
 Generator loss function:
𝐺𝑙𝑜𝑠𝑠 = −log⁡(𝐷(𝐺(𝑧)))
where 𝐺 is the generator, 𝐷 is the discriminator, 𝑧
is the noise vector, and 𝑙𝑜𝑔 is the natural logarithm.
 Discriminator loss function:
𝐷𝑙𝑜𝑠𝑠 = ⁡ −log(𝐷(𝑥)) −⁡ log(1 − 𝐷(𝐺(𝑧)))
where 𝑥 is a real image from the dataset.
 DCGAN loss function:
𝐷𝐶𝐺𝐴𝑁𝑙𝑜𝑠𝑠 = 𝐺𝑙𝑜𝑠𝑠 + 𝐷𝑙𝑜𝑠𝑠
Step 3: Feature Extraction with CNN
 Load the saved generator model
 Define the CNN model
 Freeze the weights of the DCGAN generator layers
 Train the CNN on the generated images from the
DCGAN and the real images from the dataset
 Extract the features from the CNN model
 CNN loss function:
𝐶𝑁𝑁𝑙𝑜𝑠𝑠 = ⁡ −𝑌 ∗ log(𝑝) − (1 − 𝑌) ∗ log(1 − 𝑝)
where 𝑋 is the input image, 𝑌 is the corresponding
label (real or generated), and 𝑝 is the predicted
probability of the input image being real.
 Extracted features:
𝑓𝑒𝑎𝑡𝑢𝑟𝑒𝑠 = 𝐶𝑁𝑁(𝑥)
Step 4: Combining CNN and DCGAN Fig. 5. Hybrid CNN And GAN Architecture For Deepfake Image
 Concatenate the features from the CNN and the
DCGAN generator Fig. 5 overviews a process for detecting DeepFakes that
 Define a classification layer combines CNNs and DCGANs. Preprocessing involves

Authorized licensed use limited to: K K Wagh Inst of Engg Education and Research. Downloaded on August 21,2024 at 16:22:36 UTC from IEEE Xplore. Restrictions apply.
 resizing and normalizing the photos, loading the DeepFake V. DISCUSSION
dataset and splitting it into training and testing sets. After the Since DeepFake technology may be used for harmful
DCGAN's generator and discriminator models have been goals like disseminating fake news and social defamation, its
defined, the DCGAN is trained using authentic and spoofed use has skyrocketed in recent years, making it a major threat
images. The DCGAN loss function is computed utilizing the to society. Researchers have developed several deep learning
generator and discriminator models, and the trained generator strategies, such as CNN and DCGANs, to identify and
model is stored. eliminate DeepFake material. Recent research suggests that
Finally, the DCGAN generator layers' weights are frozen, combining these architectures can increase the effectiveness
and the loaded generator model is used for feature extraction of DeepFake detection. However, due to issues like
with CNN. The features are then extracted from the CNN computational complexity and low generalizability,
model, which is trained on both DCGAN-generated images DeepFake detection continues to be difficult. The necessity
and real photos from the dataset. Similarly, the CNN loss of having enough space to keep many movies used to train
function is computed by taking the input image, the label, and machine learning models and the difficulties of recognizing
the estimated probability that the input image is real. DeepFakes are also discussed. To circumvent this, we suggest
employing the concept of virtual memory. The study
As a final step, we link CNN and DCGAN by combining highlights the importance of identifying fake audio
their respective feature sets, defining a classification layer, recordings and proposes using multiple machine models to
and training the new model. The CNN needs the input image, identify machine-generated speech and pitch or modulation
the label, and the input image's veracity prediction to mismatches. Future studies should concentrate on improving
calculate the combined loss function. Finally, the model's datasets, developing detecting algorithms, and resolving the
ability to detect DeepFakes is evaluated using validation data. shortcomings of current approaches. By doing this, we can
IV. RESULTS reduce the spread of false information and fake news,
safeguard the public's confidence, and enable the moral and
A. Proposed model accuracy measures advantageous use of DeepFake technology in various fields,
including entertainment and education.
TABLE III. ACCURACY MEASURE
CONCLUSION
Model Name Accuracy Precision Recall F1 Score
CNN 0.95 0.96 0.94 0.95 This study concludes by highlighting the growing worry
GAN 0.92 0.93 0.90 0.91 of DeepFakes, which are digitally fabricated movies, audio
Ensemble
0.97 0.98 0.96 0.97 recordings, and photographs that can propagate
Model
disinformation and fake news and modify private
Table III demonstrates that out of the three models, the information. DeepFakes could be used for bad things, like
ensemble model has the highest accuracy (0.97), precision inciting violence or manipulating public opinion, or for good
(0.98), recall (0.97), and F1 score (0.97). The CNN model things, like visual effects or interactive teaching. CNN and
performs admirably with an accuracy of 0.95, while the GAN GANS are only two examples of deep learning models that
model falls just short at 0.92. Using these measures, we can detect and eliminate DeepFakes.
assess how well the proposed DeepFake detection model
works and how it stacks up against competing algorithms. The article stresses the need for more effective detection
tools to limit the damage caused by DeepFakes and maximize
The suggested model includes CNNs, GANs, and an their usefulness. Key components of constructing efficient
Ensemble Model aggregating their results. Standard metrics
detection models include transfer learning, deep learning, and
such as true positive, true negative, false positive, and false
negative rates are used to determine the model's precision, optimization strategies.
recall, and F1 score, indicating the model's accuracy. In the end, the research community must improve
While the suggested approach does well at identifying detection algorithms to guarantee the authenticity and
DeepFakes, it may not be immune to all variations of the dependability of multimedia information as DeepFakes
problem. There is ongoing innovation in the DeepFake continue to constitute a substantial danger. By doing so, we
industry, with new methods and tools being developed to can prevent the proliferation of false information and news,
produce even more convincing forgeries. To keep up with protect public confidence, and pave the way for the ethical
these advancements, developing and refining DeepFake and productive application of DeepFake across a wide range
detection systems is crucial. of fields.
REFERENCES
Ensemble Model GAN CNN
[1] S. R. B. R, P. K. Pareek, B. S, and G. G, ‘Deepfake Video Detection
System Using Deep Neural Networks’, 2023 IEEE International
F1 Score 0.97 Conference on Integrated Circuits and Communication Systems
0.91 (ICICACS), pp. 1–6, 2023.
0.95
Recall 0.96 [2] A. S and N. Thillaiarasu, ‘Investigation Of Comparison on Modified
0.9 CNN Techniques to Classify Fake Face in Deepfake Videos’, 2022 8th
0.94
0.98 International Conference on Advanced Computing and
Precision 0.93 Communication Systems (ICACCS), vol. 1, pp. 702–707, 2022.
0.96
0.97 [3] Tian, Y., Wang, J., Wu, J., & Huang, Y. (2021). Robust detection of
Accuracy 0.92 deepfake face images using texture and deep features. Signal
0.95
Processing: Image Communication, 99, 117066.
[4] S. Chandrasekaran, V. Dutt, N. Vyas and R. Kumar, "Student
Fig. 6. Visualization of accuracy measures Sentiment Analysis Using Various Machine Learning Techniques,"

Authorized licensed use limited to: K K Wagh Inst of Engg Education and Research. Downloaded on August 21,2024 at 16:22:36 UTC from IEEE Xplore. Restrictions apply.
 2023 International Conference on Artificial Intelligence and Smart CNN Deep Learning Model’, International journal of electrical and
Communication (AISC), Greater Noida, India, 2023, pp. 104-107, doi: computer engineering systems, 2023.
10.1109/AISC56616.2023.10085018. [18] Zhang, C., Zhu, X., Song, Y., Huang, Y., Zhou, X., & Zhang, X.
[5] Vo, H. N., & Luong, M. T. (2021). Deepfake detection using a multi- (2021). Deepfake detection using convolutional neural networks with
task convolutional neural network with adversarial training. feature fusion. Journal of Visual Communication and Image
Neurocomputing, 459, 401-408. Representation, 78, 103051.
[6] Rathore, S., & Aggarwal, S. (2021). Improved Deepfake Detection [19] Zhang, C., Zhu, X., Song, Y., Huang, Y., Zhou, X., & Zhang, X.
using Attention Mechanism and Spatial Pyramid Pooling. IEEE (2021). Deepfake detection using convolutional neural networks with
Transactions on Multimedia, 23, 462-471. feature fusion. Journal of Visual Communication and Image
[7] Kumar, S. A. Kumar, V. Dutt, A. K. Dubey, S. Narang. A Hybrid Representation, 78, 103051.
Secure Cloud Platform Maintenance Based on Improved Attribute- [20] V. Dutt, S. M. Sasubilli and A. E. Yerrapati, "Dynamic Information
Based Encryption Strategies, International Journal of Interactive Retrieval With Chatbots: A Review of Artificial Intelligence
Multimedia and Artificial Intelligence, (2021). Methodology," 2020 4th International Conference on Electronics,
[8] Bappy, J. H., Roy-Chowdhury, A. K., & Chan, A. B. (2021). Detecting Communication, and Aerospace Technology (ICECA), Coimbatore,
Deepfakes in the Wild. IEEE Transactions on Pattern Analysis and India, 2020, pp. 1299-1303, DOI:
Machine Intelligence, 43, 3422-3438. 10.1109/ICECA49313.2020.9297533.
[9] Sabir, M., Lee, Y., & Kim, J. (2021). Deepfake detection using [21] S. R. Swarna, A. Kumar, P. Dixit and T. V. M. Sairam, "Parkinson’s
ensemble of fine-tuned convolutional neural networks. IEEE Access, Disease Prediction using Adaptive Quantum Computing," 2021 Third
9, 30818-30828. International Conference on Intelligent Communication Technologies
and Virtual Mobile Networks (ICICV), 2021, pp. 1396-1401, doi:
[10] Aksu, H., & Duran, B. (2020). A comprehensive study on deepfake 10.1109/ICICV50876.2021.9388628
videos: History, detection methods, and challenges. Journal of Ambient
Intelligence and Humanized Computing, 11, 257-276. [22] G. Sasubilli and A. Kumar, "Machine Learning and Big Data
Implementation on Health Care data," 4th International Conference on
[11] P. Saikia, D. Dholaria, P. Yadav, V. M. Patel, and M. Roy, ‘A Hybrid Intelligent Computing and Control Systems (ICICCS), 2020, pp. 859-
CNN-LSTM model for Video Deepfake Detection by Leveraging 864, doi: 10.1109/ICICCS48265.2020.9120906.
Optical Flow Features’, 2022 International Joint Conference on Neural
Networks (IJCNN), pp. 1–7, 2022. [23] S. M. Sasubilli, A. Kumar and V. Dutt, "Machine Learning
Implementation on Medical Domain to Identify Disease Insights using
[12] P. Saikia, D. Dholaria, P. Yadav, V. M. Patel, and M. Roy, ‘A Hybrid TMS," 2020 International Conference on Advances in Computing and
CNN-LSTM model for Video Deepfake Detection by Leveraging Communication Engineering (ICACCE), 2020, pp. 1-4, doi:
Optical Flow Features’, 2022 International Joint Conference on Neural 10.1109/ICACCE49060.2020.9154960.
Networks (IJCNN), pp. 1–7, 2022.
[24] S. M. Sasubilli, A. Kumar and V. Dutt, "Improving Health Care by
[13] L. Chen, Y. Zhang, Y. Song, L. Liu, and J. Wang, ‘Self-supervised Help of Internet of Things and Bigdata Analytics and Cloud
Learning of Adversarial Example: Towards Good Generalizations for Computing," 2020 International Conference on Advances in
Deepfake Detection’, 2022 IEEE/CVF Conference on Computer Computing and Communication Engineering (ICACCE), 2020, pp. 1-
Vision and Pattern Recognition (CVPR), pp. 18689–18698, 2022. 4, doi: 10.1109/ICACCE49060.2020.9155042.
[14] S. R. A. Ahmed, E. Sonuç, M. R. A. Ahmed, and A. D. Duru, ‘Analysis
[25] P. K. Kotturu and A. Kumar, "Comparative Study on Machine
Survey on Deepfake detection and Recognition with Convolutional
Learning models for Early Diagnose of Alzheimer's Disease: Multi
Neural Networks’, 2022 International Congress on Human-Computer
Correlation Method," 2020 5th International Conference on
Interaction, Optimization and Robotic Applications (HORA), pp. 1–7,
Communication and Electronics Systems (ICCES), 2020, pp. 778-783,
2022.
doi: 10.1109/ICCES48766.2020.9137872
[15] A. Malik, M. Kuribayashi, S. M. Abdullahi, and A. N. Khan,
[26] S. R. Swarna, A. Kumar, P. Dixit and T. V. M. Sairam, "Parkinson’s
‘DeepFake Detection for Human Face Images and Videos: A Survey’,
Disease Prediction using Adaptive Quantum Computing," 2021 Third
IEEE Access, vol. 10, pp. 18757–18775, 2022.
International Conference on Intelligent Communication Technologies
[16] Z. Gu, Y. Chen, T. Yao, S. Ding, J. Li, and L. Ma, ‘Delving into the and Virtual Mobile Networks (ICICV), 2021, pp. 1396-1401, doi:
Local: Dynamic Inconsistency Learning for DeepFake Video 10.1109/ICICV50876.2021.9388628.
Detection’, in AAAI Conference on Artificial Intelligence, 2022.
[17] S. T. Ikram, P. V, S. Chambial, D. Sood, and A. V, ‘A Performance
Enhancement of Deepfake Video Detection through the use of a Hybrid

Authorized licensed use limited to: K K Wagh Inst of Engg Education and Research. Downloaded on August 21,2024 at 16:22:36 UTC from IEEE Xplore. Restrictions apply.