Attack-Agnostic Deep Face Anti-Spoofing

The task of face anti-spoofing (FAS) is to determine whether the captured face from a face recognition system is live or fake. Current methods which are trained with existing fake faces ignore the generalization and perform poorly on unseen attacks. To tackle this problem, a novel Attack-agnostic Face Anti-spoofing framework is proposed. Different from previous methods that can be treated as a defense system, we regard face anti-spoofing as a unified framework with the attack and defense systems and optimize the defense system against unseen attacks via adversarial training with the attack system. Concretely, the attack system consists of two modules: an Adversarial learning-based Attack Pattern Generation (Adv-APG) module and a Supervised learning-based Attack Pattern Drift (Sup-APD) module. The Adv-APG module generates a series of spoofing samples by recombining a live face with known attack patterns in a generative way. The Sup-APD module pulls the generated spoofing samples in a supervised way to an unknown domain that makes the defense system ineffective. The defense system is free to choose and compatible with our attack system. Extensive experiments are conducted by using three different defense architectures to verify that the proposed attack system can improve the performance on both seen- and unseen attacks on multiple datasets.