Revisiting face forgery detection towards generalization

Abstract

Face forgery detection aims to distinguish AI generated fake faces with real faces. With the rapid development of face forgery creation algorithms, a large number of generative models have been proposed, which gradually reduce the local distortion phenomenon or the specific frequency traces in these models. At the same time, in the process of face data compression and transmission, distortion phenomenon and specific frequency cues could be eliminated, which brings severe challenges to the performance and generalization ability of face forgery detection. To promote the progress on face forgery detection research towards generalization, we present the first comprehensive overview and in-depth analysis of the generalizable face forgery detection methods. We categorize the target of generalizable face forgery detection into the robustness on novel and unknown forged images, and robustness on damaged low-quality images. We discuss representative generalization strategies including the aspects of data augmentation, multi-source learning, fingerprints detection, feature enhancement, temporal analysis, vision-language detection. We summarize the widely used datasets and the generalization performance of state-of-the-art methods in terms of robustness to novel unknown forgery as well as damaged quality forgery types. Finally, we discuss under-investigated open issues on face forgery detection towards generalization in six directions, including building a new generation of datasets, extracting strong forgery cues, considering identity features in face forgery detection, security and fairness of forgery detectors, the potential of large models in forgery detection and test-time adaptation. Our revisit of face forgery detection towards generalization will help promote the research and application of face forgery detection on real-world unconstrained conditions in the future.