Spatiotemporal Consistency-Based Deep Forgery Detection

International School, Beijing University of Posts and Telecommunications, Beijing, China

^* Corresponding author’s email: This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

With the rapid development of generative artificial intelligence (AIGC) and deep learning technology, deep forgery technology based on generated adversarial network (GAN) and diffusion model can generate extremely realistic facial videos and images, posing a serious threat to the social trust system and information security. Existing research shows that fake videos often show subtle inconsistencies in space-time dimensions, such as light changes between frames, texture jitter and abnormal motion blur. These “space-time inconsistency” have become an important basis for identifying deep forgery. This article summarizes the deep forgery detection method based on the characteristics of space-time consistency by systematically sorting out existing research. First, explain the role of space-time consistency characteristics in video forgery detection; secondly, review the typical detection framework from the three dimensions of time domain, airspace and depth, including abnormal detection based on time-domain light flow, feature fusion methods based on spatial attention, and anti-deception models combining gradient and depth estimation; then analyze commonly used Data sets and evaluation indicators; finally summarize the shortcomings of existing research and look forward to the future direction, such as multimodal fusion and cross-domain generalization. This article aims to provide systematic research reference and development ideas for deep forgery video detection.