Deepfake Face Detection from GAN to Diffusion Model
TL;DR
Analysis of image-level face forgery detection from three perspectives shows that cross-domain performance is closely related to the physical universality of the detection cues, and methods based on physical constraints exhibit more stable generalization capabilities.
Abstract
. As Generative Adversarial Networks (GANs) evolve towards diffusion models, deeply forged face images exhibit increasingly high similarity to real images in terms of pixel-level statistical characteristics. This causes detection methods based on low-level texture artifacts or local statistical distortions to fail in cross-model scenarios. This paper explores image-level face forgery detection from three perspectives: datasets, spatial domain methods, and frequency domain methods. Spatial domain methods have evolved from microscopic trace extraction and attention-based modeling to physical logic verification based on visual language models, while frequency domain methods have progressed from amplitude-based statistical analysis to phase spectrum consistency modeling. Analysis shows that cross-domain performance is closely related to the physical universality of the detection cues. Methods based on physical constraints exhibit more stable generalization capabilities, while those methods that heavily rely on the statistical characteristics of the training set experience significant performance degradation under distribution shifts.