Dual fusion paired environmental background and face region for face anti-spoofing

Abstract

Face anti-spoofing is a key phase in the face recognition process, where threats come from various deception attacks. Previously, traditional methods and deep supervised learning methods were shown to be effective in face anti-spoofing, but most previous work only focused on a single application scenario, ignoring the importance of face anti-spoofing methods for generalization ability across different applications scenarios. As a result, we propose a new face anti-spoofing method based on misleading attack information found in the face area and maybe in the environmental backdrop. The convolutional neural network extracts deception attack information from the global picture, while the local feature descriptor extracts deception attack information from the face area. The dual-cue fusion method efficiently mitigates detector performance loss caused by changes in the detection backdrop. We conduct several trials using CelebA-Spoof, WMCA, and 3DMAD datasets to demonstrate the efficiency of our technique. The findings reveal that our solution is capable of dealing with the majority of assaults and has a high degree of generality for various application scenarios.