Dual-Stream Dual-Backbone Architecture for HyperSpectral Face Anti-Spoofing

Due to limitations in detecting sophisticated Presentation Attacks using existing techniques, FAS research utilizing hyperspectral images is gaining attention. In this paper, we propose a Dual-Stream Dual-Backbone Network (DSDBNet) for detecting hyperspectral spoofed facial images. DSDBNet processes hyperspectral facial images through two streams, with each stream using a backbone model of identical structure but not sharing parameters. DSDBNet fuses features extracted from the backbone models of each stream to perform final classification, achieving an 11% performance improvement compared to the parameter-sharing SiameseNet. We expect that the proposed DSDBNet will significantly improve performance in hyperspectral spoofed face detection.