CG-FAS: Cross-label Generative Augmentation for Face Anti-Spoofing

Xing Liu, Anyang Su, Minghui Wu, Zitong Yu, Kangle Wu, Da An, Jie Hao, Mengzhen Xu, Chenxu Zhao, Zhen Lei

Abstract: Face Anti-Spoofing (FAS) is essential to secure face recognition systems from various physical attacks. A sufficient and diverse training set helps to build robust FAS models. To exploit the potential of FAS datasets, we propose to generate high-quality data including live and diverse presentation attacks (PAs) faces, for data augmentation during the model training stage. Our method is called Cross-label Generative augmentation for Face Anti-Spoofing (CG-FAS), which could convert a live face into a 3D high-fidelity mask, replay, print, or other extra physical PAs. Correspondingly, CG-FAS can also restore a specific physical presentation attack into a live face. This function is realized by innovatively building an Interchange Bridge matrix, which stores disentangled spoof clues between PAs and live faces. To verify the effects of these generated data, we utilize them as augmentation data and conduct experiments on several typical FAS benchmarks. Extensive experimental results demonstrate the superior performance gain with CG-FAS for off-the-shelf data-driven FAS models. We hope the CG-FAS can shine a light on the deep FAS community to alleviate the data-hungry issue. The code will be released soon at: https://github.com/liuxingwt/CG-FAS.