Deepfake detection with domain generalization and mask-guided supervision

Abstract

Most existing deepfake (video face forgery) detectors work well in intra-dataset testing, but their performance degrades severely in cross-dataset testing. Cross-dataset generalization remains a major challenge. Since domain generalization (DG) aims to learn domain-invariant features while suppressing domain specific features, we propose a DG framework for improving face forgery detection in this study. Our detector consists of two modules. The first module learns both spatial and spectral features from frame images. The second one learns high-level feature patterns from the outputs of the first module, and constructs the classification features with the help of face mask-guided supervision. The classification result is fine-tuned by a confidence-based correction mechanism. The DG framework is realized through a bi-level optimization process. Extensive experiments demonstrate that our detector works effectively in both intra- and cross-dataset testing. Compared with 8 typical methods, it has the best overall performance and the highest robustness against common perturbations.