Face Forgery Detection by Multi-dimensional Image Decomposition

Tianshuo Zhang, Xiangyu Zhu, Feng Pan, Ke Xiang, Zhen Lei

Abstract: With the development of face manipulation techniques and the potential harms of fake videos and images to public safety, the field of face forgery detection has received a lot of attention. However, existing face forgery detection methods mostly rely on specific datasets, and the generalization performance appears to be degraded when testing across datasets. In this paper, we use multi-dimensional image decomposition, i.e., spatial-frequency decomposition and computer graphics decomposition, to provide additional forgery clues. On the one hand, we decompose the face image into three components: high-frequency component, medium frequency component, and low-frequency component using spatial-frequency decomposition; on the other hand, we use computer graphics algorithms to simulate the face generation process, separate the 3D model and decompose the face image into five components. For the decomposed components, we introduce a three-stream neural network to fuse multi-modality information. Extensive experiments have shown that our method achieves state-of-the-art performance.