Face Anti-spoofing with Unknown Attacks: A Comprehensive Feature Extraction and Representation Perspective

Face anti-spoofing aims at detecting whether the input is a real photo of a user (living) or a fake (spoofing) image. As new types of attacks keep emerging, the detection of unknown attacks, known as Zero-Shot Face Anti-spoofing (ZSFA), has become increasingly important in both academia and industry. Existing ZSFA methods mainly focus on extracting discriminative features between spoofing and living faces. However, the nature of the spoofing faces is to trick anti-spoofing systems by mimicking the livings, so the deceptive features between the known attacks and the livings, which have been ignored by existing ZSFA methods, are essential to comprehensively represent the livings. Therefore, existing ZSFA models are incapable of learning the complete representations of living faces and thus falling short on effectively detecting newly-emerged attacks. To address this issue, we propose a novel feature extraction framework that can capture both the deceptive and discriminative features between living and existing spoofing faces. This framework is composed of a learnable masking mechanism and a two-against-all training scheme. To address the subsequent invalidation issue of categorical functions and dominance disequilibrium issue among different dimensions of features after importing deceptive features, we employ a newly modified semantic autoencoder to represent all extracted features to a semantic space to equilibrate the dominance of each feature dimension. As a result, our method simultaneously achieves a feasible detection on unknown attacks and a comparably accurate detection on known spoofing. Experimental results confirm the superiority and effectiveness of our proposed method in identifying the livings with the interference of both known and unknown spoofing types.