Analysis of Generative Data Augmentation for Face Antispoofing

Analysis of Generative Data Augmentation for Face Antispoofing

연구 분야: Strategies

학회: International Conference on Pattern Recognition Applications and Methods

As technology advances, criminals continually find innovative ways to gain unauthorised access, increasing face spoofing challenges for face recognition systems. This demands the development of robust presentation attack detection methods. While traditional face antispoofing techniques relied on human-engineered features, they often lacked optimal representation capacity, creating a void that deep learning has begun to address in recent times. Nonetheless, these deep learning strategies still demand enhancement, particularly in uncontrolled environments. In this study, we employ generative models for data augmentation to boost the face antispoofing efficacy of a vision transformer. We also introduce an unsupervised keyframe selection process to yield superior candidate samples. Comprehensive benchmarks against recent models reveal that our augmentation methods significantly bolster the baseline performance on the CASIA-FASD dataset and deliver state-of-the-art results on the Spoof in the Wild database for protocols 2 and 3.