Learning to Discover Forgery Cues for Face Forgery Detection

Locating manipulation maps, i.e., pixel-level annotation of forgery cues, is\ncrucial for providing interpretable detection results in face forgery\ndetection. Related learning objects have also been widely adopted as auxiliary\ntasks to improve the classification performance of detectors whereas they\nrequire comparisons between paired real and forged faces to obtain manipulation\nmaps as supervision. This requirement restricts their applicability to unpaired\nfaces and contradicts real-world scenarios. Moreover, the used comparison\nmethods annotate all changed pixels, including noise introduced by compression\nand upsampling. Using such maps as supervision hinders the learning of\nexploitable cues and makes models prone to overfitting. To address these\nissues, we introduce a weakly supervised model in this paper, named Forgery Cue\nDiscovery (FoCus), to locate forgery cues in unpaired faces. Unlike some\ndetectors that claim to locate forged regions in attention maps, FoCus is\ndesigned to sidestep their shortcomings of capturing partial and inaccurate\nforgery cues. Specifically, we propose a classification attentive regions\nproposal module to locate forgery cues during classification and a\ncomplementary learning module to facilitate the learning of richer cues. The\nproduced manipulation maps can serve as better supervision to enhance face\nforgery detectors. Visualization of the manipulation maps of the proposed FoCus\nexhibits superior interpretability and robustness compared to existing methods.\nExperiments on five datasets and four multi-task models demonstrate the\neffectiveness of FoCus in both in-dataset and cross-dataset evaluations.\n