An Improved GAN-based Depth Estimation Network for Face Anti-Spoofing.

Face spoofing is a serious threat to the security of face recognition and face anti-spoofing (FAS) has become one of the popular research directions recently. The pipeline of modern face anti-spoofing methods is to train deep neural networks to distinguish live and spoof faces using various auxiliary information. Depth maps are the most commonly used auxiliary features due to their computational simplicity and ease of discrimination. However, existing methods do not generalize well in complex environments and against unknown attacks. In this paper, we propose a novel FAS method using generative adversarial network (GAN). We train a GAN for generating depth maps of faces. The network uses the architecture of Wasserstein GAN and adds an attention module to learn the contribution of different regions to the deep graph. The classifier is trained using a latent variable containing depth information in the network. Experiments demonstrate that our approach can better generalize complex external conditions such as illumination and background. Among all results, we gain the best result of 1.17% (EER) on CASIA-FASD and 21.57% (HTER) on cross testing between Replay-Attack and CASIA-FASD.