Generalized face anti-spoofing with data synthesis based on generative model

Abstract

Face anti-spoofing (FAS) is a critical task in ensuring the security and reliability of facial recognition systems. To improve the generalizability to untrained attacked types of face image, the FLIP framework (Face anti-Spoofing with Language guidance for Cross-domain applications) introduces a novel approach by incorporating language guidance to address the challenges posed by cross-domain spoofing. FLIP utilizes textual descriptions to assist in distinguishing between real and spoofed faces, improving the model’s robustness across different spoofing techniques and domains. In this dissertation, FLIP approach is extended by assigning each image with more specific prompts in perspective of typical view or judgement of distinguishing real and spoofed face. This dissertation mainly follows the multimodal contrastive learning strategy and also conduct experiments to see how descriptions affect the results. The effect of how the prompts affecting results is recorded and compared. The comparisons between FLIP and our approach demonstrate that more specific textual guidance yields higher accuracy, and the ablation study reveals the increasing the number of prompts does not necessarily enhance performance.