Supervisors: Dr. Suman Saha, Dr. Danda Pani Paudel, and Prof. Luc van Gool
Nowadays, due to the universal utilization of mobile and computing devices, a demand for robust authentication systems has driven the interest of many researchers. Face recognition systems, specifically, has become the target of various presentation attacks such as 3D plastic masks and video replays. Face anti-spoofing (FAS) techniques, are developed to create a robust system against such attacks. Thanks to the recent development of deep neural networks, the overall accuracy of face anti-spoofing methods has been dramatically increased. Most state-of-the-art anti-spoofing methods focuses on detecting artifacts from training samples such as images and videos. However, in most cases people use training datasets, which are always extracted in an indoor scene with little variation among backgrounds, illumination conditions, camera resolutions, spoof materials, etc. As a consequence of such training strategy, most of the face anti-spoofing architectures have poor generalization ability, which means they are likely to fail when a new sample came. Our motivation is to improve the overall training strategy of face anti-spoofing systems so as to increase their ability to generalize in new scenarios. We propose to a domain generalization face anti-spoofing network, which uses gradient reversal layers on top of a spatial-temporal deep neural network. Extensive experimental analysis have shown that our proposed network structure can significantly outperform the state-of-the-art face anti-spoofing method as well as give out understandable visualization results.