Deepfake video detection methods, approaches, and challenges

Abstract

Deepfake technology creates highly realistic manipulated videos using deep learning models, which makes distinguishing between authentic and fake content extremely difficult. This technology can negatively affect society by breaching privacy and spreading misinformation. This paper presents a comprehensive survey of the recent deepfake video detection approaches and methods. Each deepfake video method is analyzed according to its ability to generalize diverse deepfake fabrication techniques and real-world scenes. We reviewed around 103 articles which eventually shrunk down to 73 based on the screening criteria like abstract/title/irrelevant focus/duplication. The study primarily covers audio-based, visual-based, and multi-modal detection methods. Also, it discusses the usage of Convolutional Neural Networks (CNNs), frequency-domain analysis, and audio-visual synchronization in deepfake video detection and evaluates the strengths and shortcomings of these techniques. Moreover, the study explores major issues such as low resolution, video compression, and adversarial attacks, which prove to be a barrier to making deepfake video detection processes robust. By connecting findings from numerous studies, this research draws attention to the development of standard benchmarking SOPs and multi-modal detection techniques to improve detection performance.