Patch Diffusion: A General Module for Face Manipulation Detection

Detection of manipulated face images has attracted a lot interest recently. Various schemes have been proposed to tackle this challenging problem, where the patch-based approaches are shown be promising. However, existing tend treat different patches equally, which do not fully exploit patch discrepancy for effective feature learning. In paper, we propose Patch Diffusion (PD) module can integrated into manipulation detection networks boost performance. The PD consists Discrepancy Feature Learning (DPFL) and Attention-Aware Message Passing (AMP). DPFL effectively learns features by newly designed Pairwise Loss (PPLoss), takes both importance correlations consideration. AMP diffuses through attention-aware message passing in graph network, attentions explicitly computed based on learnt DPFL. We integrate our four recent networks, carry out experiments popular datasets. results demonstrate that is able performance detection.