PROJECT TITLE :

Meta-Teacher For Face Anti-Spoofing

ABSTRACT:

Face anti-spoofing, also known as FAS, protects Face Recognition systems from fraudulent presentation attempts (PAs). The currently available FAS methods typically supervise PA detectors with labels that are either handcrafted binary or pixel-wise. On the other hand, handcrafted labels might not be the best way to supervise PA detectors as they learn sufficient and intrinsic spoofing cues. We propose a novel Meta-Teacher FAS (MT-FAS) method to train a meta-teacher for more effectively supervising PA detectors in place of the handcrafted labels. This will allow for greater accuracy. The meta-teacher is trained in a bi-level optimization manner to learn the ability to supervise the PA detectors as they learn rich spoofing cues. This training is done in order for the meta-teacher to acquire this capability. The bi-level optimization is comprised of two essential elements, namely: 1) a lower-level training in which the meta-teacher supervises the learning process of the detector on the training set; and 2) a higher-level training in which the meta-teaching teacher's performance is optimized by minimizing the detector's validation loss. Both of these trainings are performed in sequence. Our meta-teacher is specifically trained to improve its ability to teach the detector (the student), in contrast to existing teachers, who are trained to achieve high levels of accuracy at the expense of their ability to instruct. This is one of the primary ways in which our meta-teacher stands out from other teacher-student models. Extensive experiments on five FAS benchmarks show that the proposed MT-FAS, with its trained meta-teacher, 1) provides better-suited supervision than both handcrafted labels and existing teacher-student models, and 2) significantly improves the performances of PA detectors. Both of these claims are supported by evidence from the aforementioned experiments.