Many existing face anti-spoofing (FAS) methods focus on modeling the decision
boundaries for some predefined spoof types. However, the diversity of the spoof
samples including the unknown ones hinders the effective decision boundary
modeling and leads to weak generalization capability.