Real Vs Fake: Detecting Deep Fake Videos Based ON Temporal changes in the Optical Flow between Frames

Abstract

This research analyses different methods to detect deepfakes and proposes a method that is novel and computationally competent. Main methods examined range through deep learning-based methods that use CNNs and GAN analysis, machine learning, and statistical methods. It also talks about the important issue of fairness in deepfake detection, with the goal of reducing biases across different demographic groups i.e. the proposed method is independent of features that may render it racist. This study introduces novel features designed to yield results comparable to complex and computationally intensive existing methodologies. The choice of new features was based on the idea that the way facial movements change over time will be different in deepfakes than in the original video. The frequency spectrum of temporal variations makes the difference clear. It also talks about the important issue of fairness in deepfake detection, with the goal of reducing biases across different demographic groups. It also looks at how detection methods can be used in media forensics workflows, even testing the usefulness of hand-crafted features alongside learnt representations. The need for robust and efficient detection in the face of evolving deepfake technologies and the importance of interpretable and fair solutions is felt more than ever before.