Date of Publication :12th July 2024

Abstract:Deepfakes are the latest - and fast-developing - form of attack on digital video and audio. They exploit the recent breakthroughs in machine learning technology, specifically Generative Adversarial Networks (GANs), to produce extremely realistic fake video. Deepfakes can swap faces or even synthesize entire facial gestures with a high-level of craft that is hard to distinguish between the real and generated content. With the rising quality of deepfakes, robust video forensic methods are necessary to detect and verify their presence. Human analysis and basic heuristic methods have failed against well crafted deepfakes. Consequently, recent studies suggest machine learning and computer vision approaches for their detection. Some popular spatial detection methods include Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks and hybrid spatial-temporal models that search for subtle inconsistencies in the video. Recent datasets like FaceForensics++, Celeb-DF and the Deep Fake Detection Challenge (DFDC) have been proposed for training and testing detection models. Optical flow based techniques have also been integrated with detection models to search for inconsistencies in the motion fields of successive frames. This survey discusses the state-of-the-art techniques for the generation and detection of deepfakes and encourages the development of video forensics to check the authenticity of a given content.

Reference :