Dual-stream correlation exploration for face anti-Spoofing

Abstract

Face anti-spoofing (FAS) is an important technology to ensure the security of face recognition system. Previous methods generally focus on the representation of the spoof patterns. However, the hidden correlations between living and spoofing faces are ignored, making the generation and discrimination of face anti-spoofing less effective. In this paper, we propose a novel Dual-Stream Correlation Exploration method (DSCE) to simultaneously model the correlation between content and liveness features for the performance improvement. Specifically, DSCE devises two novel modules: the spoof cue generation module and the source face reconstruction module, at two streams respectively. The former one introduces pseudo negative feature to extend the diversity of attack types, and adopts a metric learning strategy to learn the correlation among liveness features. The latter one integrates liveness and content features to explores the potential relationship between the both features through source face reconstruction. Last, DSCE adaptively combines spoof cues and reconstructed faces to comprehensively consider the importance of different correlations for face anti-spoofing. Comprehensive experimental results on both intra-dataset testing and cross-dataset testing clearly demonstrate the high discrimination and generalization of DSCE.