Face Manipulation Detection via Auxiliary Supervision

The rapid progress of face manipulation technology has attracted people’s attention. At present, a reliable edit detection algorithm is urgently needed to identify real and fake faces to ensure social credibility. Previous deep learning approaches formulate face manipulation detection as a binary classification problem. Many works struggle to focus on specific artifacts and generalize poorly. In this paper, we design reasonable auxiliary supervision to guide the network to learn discriminative and generalizable cues. A multi-scale framework is proposed to estimate the manipulation probability with texture map and blending boundary as auxiliary supervisions. These supervisions will guide the network to focus on the underlying texture information and blending boundary, making the learned features more generalized. Experiments on FaceForensics and FaceForensics++ datasets have demonstrated the effectiveness and generalization of our method.