Information Adversarial Disentanglement for Face Swapping

Tingting Chen, Yang Yu, Rongrong Ni, Yao Zhao

Abstract: Face swapping can provide data support for face forgery detection, which is a very significant topic in forensics. It is the task of converting the source identity to the target face while preserving target attributes, thus disentangling identity and identity-unrelated (i.e., attribute) features is still a challenging task. In this work, we focus on intra-class (i.e. identities) and inter-class (i.e. identity and attribute) relationships to comprehensively decouple identity and attribute features in an adversarial way for face swapping. The whole network includes Identity-Attribute Adversary (IAA) module, Identity Reconstruction (IR) module and Re-Feeding module. Specifically, for the inter-class relationship, we first propose the IAA module to initially extract independent identity and attribute features. Besides, the Re-Feeding module re-disentangles the generated images and reconstructs original images to further confirm the complete disentanglement of the inter-class information. Finally, for the intra-class relationship, we adopt the IR module based on the same identity image pairs to learn the consistent identity feature without being influenced by attributes. Extensive experiments and comparisons to the existing state-of-the-art face swapping methods demonstrate the effectiveness of our framework.