Abstract

Detection of deepfake videos faces serious generalization problem in real world application scenarios. Existing robust deepfake detection methods can only works on single frame image but not continuous frame videos. In this paper, we propose a robust deepfake video detection method based on continuous frame face-swapping. We design our face-swapping dataset with Delaunay triangulation and piecewise affine transform to achieve continuous frame face-swapping. We design a feature enhancement module with facial and background information covered to make the method focus on the mask fusion zone. We build our detection model with Efficient Net to extract intra-frame fusion feature and LSTM to extract inter-frame time feature. Cross-domain experiments show that our method achieves better detection AUC than existing methods, which proves our method is robust because of generalization.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call