K.S.

INSTITUTE OF TECHNOLOGY
Department of Electronics and Communication Engineering
(2024-25)
Technical Seminar
on
Hybrid Deep Learning Model Based on GAN and RESNET for
Detecting Fake Faces
Tejashree N
1KS21EC105

Dept. of ECE,KSIT 1
 CONTENTS
• Introduction
• Problem Statement
• Literature survey
• Objectives
• Methodology
• Results
• Conclusion
• References

Dept of ECE,KSIT 2
 INTRODUCTION
• Deepfake is an artifiacial intelligence generated media .

• The rise of deepfake technology has led to an increase in the creation of fake faces, posing
significant challenges in various fields, including security and privacy.

• The paper introduces a hybrid deep learning model that combines Generative
Adversarial Networks (GANs) and Residual Networks (ResNet) to detect fake faces
with higher accuracy.

• The aim is to develop a effective system for fake face detection, which is essential for
combating online misinformation, fraud, and privacy violations.

Dept of ECE,KSIT 3
 LITERATURE SURVEY
Paper Methodology Advantage Disadvantage
No
1 Graph Neural Network GNNs enhance deepfake detection and This graph based approach
(GNN) improves detection accuracy may not scale well for large
video datasets

2 ResNet-50 Enhances classification performance by It may face scalability

leveraging the strengths of deep learning challenges
techniques

3 patch-wise deep learning Improves deepfake detection accuracy by The patch-wise approach may
model for multiclass breaking down faces into smaller patches lead to increased complexity
deepfake face detection and might struggle with highly
manipulated deepfakes.

deep learning approach for Provides a dual approach that not only The dual-purpose design might
4 both detecting and detects deepfakes but also generates them increase model complexity and
generating deepfakes demands, potentially limiting
its scalability

Dept of ECE,KSIT 4
 PROBLEM STATEMENT

Fake face generation technologies, such as GANs, have advanced significantly,

making it difficult to distinguish between real and fake faces. Traditional detection
methods are often insufficient, and there is a need for advanced machine learning
models that can handle the complexities of detecting fake images . The paper
identifies gaps in current methods and proposes an innovative solution by
combining GANs and ResNet to improve detection accuracy.

Dept of ECE,KSIT 5
 OBJECTIVES
o To develop a hybrid model that combines GAN for fake face generation
and ResNet for feature extraction and classification.
o Improving the accuracy and robustness of fake face detection.
o Evaluating the performance of the model on multiple datasets and
comparing it with existing methods.
o Finding solutions to the challenges of identifying deepfakes

Dept of ECE,KSIT 6
METHODOLOGY

Figure : ResNet Architecture

Dept of ECE,KSIT 7
 Figure : Architecture of GAN

Dept of ECE,KSIT 8
Figure : Overall Architecture of the Proposed Model

Dept of ECE,KSIT 9
 RESULT
Precision Recall F1-score Accuracy ROC AUC score
0.7916 0.8824 0.8345 0.8298 0.825

Table : Results of the proposed model

Precision – Measures the fraction of positive predictions that are

actually positive
Recall - Ability to capture most of the fake faces
F1-score - Balanced evaluation between precision and recall
Accuracy - Measures the overall fraction of correct predictions
ROC AUC - the ROC AUC score would assess how well the model
distinguishes between real and fake faces.

Dept of ECE,KSIT 10
Figure : False positive rate vs true positive rate for the proposed
model.

Dept of ECE,KSIT 11
ROC Curve: This is a graphical representation of the trade-off between True Positive Rate (TPR) or
Sensitivity (Recall) and False Positive Rate (FPR) across different thresholds. The ROC curve plots
TPR on the y-axis and FPR on the x-axis for various threshold values of the model’s predicted
probabilities.
 True Positive Rate (TPR), also known as Recall or Sensitivity, is calculated as:

where:
o TP (True Positive): The number of positive instances correctly classified as positive.
o FN (False Negative): The number of positive instances incorrectly classified as negative.
 False Positive Rate (FPR) is calculated as:

where:
o FP (False Positive): The number of negative instances incorrectly classified as positive.
o TN (True Negative): The number of negative instances correctly classified as negative.

Dept of ECE,KSIT 12
AUC (Area Under the Curve): This is a numerical value representing the area under the ROC curve.
A common method to calculate the AUC is:

The AUC score ranges from 0 to 1:

o A score of 1 indicates perfect classification (the model is able to perfectly distinguish between positive
and negative classes).
o A score of 0.5 indicates no discrimination ability, meaning the model is essentially making random
predictions.
o A score closer to 1 means better model performance.

Dept of ECE,KSIT 13
 CONCLUSION

 The paper concludes that the hybrid deep learning model combining GAN and ResNet
offers a powerful solution for fake face detection.
 The proposed approach outperforms traditional techniques and offers a more robust
method for dealing with advanced fake face generation technologies like GANs.
 The model shows promise for use in applications such as social media monitoring,
online security, and fraud detection.
 Future work may focus on improving model performance in real-time applications,
enhancing its ability to detect more complex manipulations, and integrating it with
other technologies.
Dept of ECE,KSIT 14
 REFERENCES
[1] M.M.El-Gayar,M.Abouhawwash,S.S.Askar,andS.Sweidan,‘‘Anovel approach for detecting deep fake videos
using graph neural network,’’ J. Big Data, vol. 11, no. 1, p. 22, Feb. 2024
[2] E. Hassan, M. S. Hossain, A. Saber, S. Elmougy, A. Ghoneim, and G. Muhammad, ‘‘A quantum convolutional
network and ResNet (50) based classification architecture for the MNIST medical dataset,’’ Biomed. Signal
Process. Control, vol. 87, Jan. 2024, Art. no. 105560.
[3] M. A. Arshed, S. Mumtaz, M. Ibrahim, C. Dewi, M. Tanveer, and S. Ahmed, ‘‘Multiclass AI-generated deepfake
face detection using patch wise deep learning model,’’ Computers, vol. 13, no. 1, p. 31, Jan. 2024.
[4] D. AbdElminaam, N. Sherif, Z. Ayman, M. Mohamed, and M. Hazem, ‘‘DeepFakeDG: A deep learning approach
for deep fake detection and generation,’’ J. Comput. Commun., vol. 2, no. 2, pp. 31–37, Jul. 2023.
[5] J. Parmar, S. Chouhan, V. Raychoudhury, and S. Rathore, ‘‘Open-world machine learning: Applications,
challenges, and opportunities,’’ ACM Comput. Surv., vol. 55, no. 10, pp. 1–37, Oct. 2023.
[6] Y.Zhu, C.Zhang, J. Gao, X. Sun, Z. Rui, and X. Zhou, ‘‘High-compressed deepfake video detection with
contrastive spatiotemporal distillation,’’ Neurocomputing, vol. 565, Jan. 2024, Art. no. 126872.

Dept of ECE,KSIT 15
[7] N. A. S. Eldien, R. E. Ali, and F. A. Moussa, ‘‘Real and fake face detection: A comprehensive evaluation of
machine learning and deep learning techniques for improved performance,’’ in IEEE MTT-S Int. Microw.
Symp. Dig., Jul. 2023, pp. 315–320.
[8] S. Duan, W. Pan, Y. Leng, and X. Zhang, ‘‘Two ResNet mini archi tectures for aircraft wake vortex
identification,’’ IEEE Access, vol. 11, pp. 20515–20523, 2023.
[9] C. Chen and B. Li, ‘‘An interpretable channelwise attention mechanism based on asymmetric and skewed
Gaussian distribution,’’ Pattern Recog nit., vol. 139, Jul. 2023, Art. no. 109467
[10] H. M. T. Khushi, T. Masood, A. Jaffar, S. Akram, and S. M. Bhatti, ‘‘Performance analysis of state-of-the-art
CNN architectures for brain tumour detection,’’ Int. J. Imag. Syst. Technol., vol. 34, no. 1, Jan. 2024, Art. no.
e22949.

Dept of ECE,KSIT 16
THANK YOU

Dept of ECE,KSIT 17