Domain Adaptation Based Person-Specific Face Anti-spoofing Using Color Texture Features

Abstract

Face anti-spoofing technology is indispensable for the face recognition system, which is vulnerable to malicious spoofing attacks such as printed attacks and replayed video attacks. In this paper, we focus on more challenging cross-database generalization, which can reflect the performance of face anti-spoofing methods in the practical application scenario. In general, the discrepancy between the sample distribution of different databases can be reduced by domain adaptation algorithms. For texture-based face anti-spoofing methods, the identity of subjects can significantly influence the effectiveness of domain adaptation. Thus, in this paper, we propose a domain adaptation based person-specific face anti-spoofing method to improve cross-database generalization. The feature used here is the local directional number pattern (LDN) extracted from HSV and YCbCr color spaces. For each target subject, we synthesize virtual fake samples for training using subject domain adaptation method. To further improve the generalization performance, we use the domain adaptation method to reduce the discrepancy between sample distribution of training and testing samples. To evaluate the performance of the proposed method, we perform cross-database experiments on CASIA and Replay-Attack database. Our method can realize promising generalization performance, outperforming most of the recently proposed methods.