BASN: Enriching feature representation using bipartite auxiliary supervisions for face anti-spoofing

Abstract

Face anti-spoofing is an important task to assure the security of face recognition systems. To be applicable to unconstrained real-world environments, generalization capabilities of the face anti-spoofing methods are required. In this work, we present a face anti-spoofing method with robust generalization ability to unseen environments. To achieve our goal, we suggest bipartite auxiliary supervision to properly guide networks to learn generalizable features. We propose a bipartite auxiliary supervision network (BASN) that comprehensively utilizes the suggested supervision to accurately detect presentation attacks. We evaluate our method by conducting experiments on public benchmark datasets and we achieve state-of-the-art performances.