The Development of Video-Oriented Deepfake Detection Technologies

With the development of deep learning technology, the spread of video-level deepfake content has posed numerous security risks and single-modal detection has become a research focus due to its strong adaptability. This paper presents a comprehensive review of video-level single-modal deepfake detection technologies, focusing on sorting out and analyzing two core categories of detection techniques: those based on temporal consistency and those based on biometric features. It elaborates on their implementation methods, core ideas and performance characteristics. Among them, Temporal consistency-based methods include CNN-LSTM fusion, thumbnail layout learning, dual-branch neural networks and two-stream convolution-ViT frameworks, while biometric-based methods cover facial motion detection, fine-grained feature extraction, 3D biometric modeling, GAN forgery-adaptive detection and microexpression dual-branch detection. Meanwhile, this paper systematically summarizes the advantages and applicable scenarios of various methods. The findings show that the two types of technologies are complementary; despite improved performance of new methods, all face certain technical bottlenecks. This paper clarifies the core technical system and existing shortcomings of video-level single-modal deepfake detection, providing theoretical reference and guidance for the subsequent optimization, innovation and engineering implementation of related technologies.