Abstract
Face manipulation detection has been receiving a lot of attention for thereliability and security of the face images/videos. Recent studies focus onusing auxiliary information or prior knowledge to capture robust manipulationtraces, which are shown to be promising. As one of the important face features,the face depth map, which has shown to be effective in other areas such as facerecognition or face detection, is unfortunately paid little attention to inliterature for face manipulation detection. In this paper, we explore thepossibility of incorporating the face depth map as auxiliary information forrobust face manipulation detection. To this end, we first propose a Face DepthMap Transformer (FDMT) to estimate the face depth map patch by patch from anRGB face image, which is able to capture the local depth anomaly created due tomanipulation. The estimated face depth map is then considered as auxiliaryinformation to be integrated with the backbone features using a Multi-headDepth Attention (MDA) mechanism that is newly designed. We also propose anRGB-Depth Inconsistency Attention (RDIA) module to effectively capture theinter-frame inconsistency for multi-frame input. Various experimentsdemonstrate the advantage of our proposed method for face manipulationdetection.