Employing Transfer-Learning based CNN architectures to Enhance the Generalizability of Deepfake Detection

Abstract

Advancements in machine learning have given rise to technologies and methodologies that are being put to use for immoral purposes, especially after the inception of Generative Adversarial Networks in 2014. Generative Adversarial Networks are capable of synthesizing hyper-realistic fake images and even videos. Various sophisticated machine learning techniques capable of creating ultra-realistic Deep Fake videos are being used to harass, blackmail women and children, induce political instability by spreading false, malicious propaganda which could in turn lead to social, political conflicts and outbursts with dire consequences. This poses a serious threat to personal safety and also endangers national security which calls for automated ways to detect deep fake videos. This paper proposes a method to expose such fake videos by using a CNN architecture and leveraging the technique of Transfer Learning. The proposed model implements an algorithm that uses a CNN for feature extraction from every frame in a video to train a binary classifier that learns to efficiently differentiate between real and manipulated videos. The method is evaluated against an extensive set of deepfake videos gathered from various datasets.