Face Anti-spoofing via Robust Auxiliary Estimation and Discriminative Feature Learning

Face anti-spoofing is critical to applications which heavily rely on the authenticity of detected faces. Recently, auxiliary information, such as facial depth maps and rPPG signals, have been successfully included to boost the performance of face anti-spoofing. Consequently, the quality of auxiliary estimation is key to the effectiveness of live/spoof classification. In this paper, we focus on the robustness of auxiliary estimation and the discriminability of latent features. We propose to estimate the auxiliary information along with the training of live/spoof classifier in an adversarial learning framework. We include additional constraints in the contrastive loss and propose a discriminative batch-contrastive loss to learn the latent features. Both the auxiliary information and the discriminative latent features are included into the live/spoof classification. In addition, because not all the auxiliary supervisions are equally reliable, we propose an adaptive fusing strategy to fuse the estimation results from different auxiliary-supervised branches. Experimental results on several benchmark datasets show that the proposed method significantly outperforms previous methods.