A Comprehensive Study on Mitigating Synthetic Identity Threats Using Deepfake Detection Mechanisms
Shivansh Uppal, Vinayak Banga, Sakshi Neeraj, Abhishek Singhal · 2024
Synthetic Identity Threats (SIT) present one of the greatest risks that could undermine the integrity and security of digital systems. These threats adapt Deepfake Technology through generation of new faces, modification of facial attributes or reenactment of fake emotions on human faces. This study is a full documentation of how Deepfake Detection mechanisms can be modelled to neutralize these SIT's using various Deep Learning architectures. We also explored applications of Deepfake beyond malicious intent into forms that are less nefarious like entertainment purposes. We utilized Generative Adversarial Networks to create Deepfakes and experimented with conducting face swaps between two distinct photos as well to assess the quality of Deepfake Technology. For detection of Deepfakes, Convolution Neural Network has shown the highest accuracy of 89.36%, Inception ResNet attained an accuracy of 82.978% while Visual Geometry Group and EfficientNet were at a score of 82.8% and 83% respectively. It is evident that although the current detecting methods are competent in their abilities, the SIT environment continues to develop. Through analyzing the nuances of producing Deepfakes and comparing current quality detection methods, we hope to offer useful information on how scholars and administrators can better protect the internet from deceitful attacks of Deepfakes.