A Survey on Face Spoof Attack Detection Using Multimodal Machine Learning Approaches

Abstract

Face anti-spoofing (FAS) has lately received more attention because of its critical function in protecting face recognition systems against presentation assaults (PAs).  With the advent of more realistic face recognition systems of various sorts, early-stage face anti-spoofing approaches based on hand-crafted characteristics have become untrustworthy due to their limited representativeness.  With the advent of large-scale academic datasets over the last decade, deep learning-based face anti-spoofing algorithms have outperformed and dominated the area.  However, most existing evaluations in this field rely on hand-crafted features, which are now outdated and hinder progress in the face anti-spoofing research community.

To inspire future research, we give the first complete analysis of recent advances in deep learning-based face anti-spoofing systems.  This article addresses numerous innovative and significant components:  1) In addition to binary label supervision (e.g., "0" for truth versus "1" for face recognition systems), we also examine cutting-edge approaches using pixel-level supervision (e.g., pseudo depth map).  2) In addition to traditional evaluation within the dataset, we collect and analyze cutting-edge methods specifically designed for domain generalization and open-set FAS analysis; and 3) In addition to commercial RGB cameras, we summarize deep learning applications under multimodal (e.g., depth and infrared) or specialized (e.g., light field and flash) sensor categories.  We end our survey by stressing existing outstanding challenges and future opportunities.