A Feasibility Study on Infrared Imaging-Based Face Antispoofing

The face recognition technology has made rapid progress in recent years, leading to the emergence of various commercial applications. However, the face, as a biometric modality, is inherently prone to leakage, forgery, and attack. Thus, the ability of face antispoofing (FAS) is crucial to a practical face recognition system and the security issue of face recognition has gradually become the focus of research for this technology. So far, many research works on FAS concentrate on designing algorithms that rely on either hand-crafted operators or deep learning models. But most of those methods are limited to a controlled circumstance – the test dataset needs to be collected under the same or at least a very similar acquisition setting. In other words, the generalization ability is limited with those works. In view of this, we in this research turn to a quite different approach: using the infrared (IR) as the imaging means and deep learning to build the algorithm. Experimental results demonstrate that IR is favorable than visible light on three different multispectral face spoofing datasets that we prepare by ourselves. It is also observed that deep learning-based algorithms are performing better than traditional hand-crafted methods, with Vision Transformer being the best, in both same-dataset testing and cross-dataset testing.