Face Presentation Attack Detection Based on Exclusivity Regularized Attention Maps

Abstract

With the wide spread of face recognition systems (FRSs) in our daily life, the security problem of FRSs, particularly presentation attack (PA) with printed photos or recorded videos, is becoming more and more serious. Inspired by the finding of prior studies that different regions on faces seem to have different contributions to the detection of PA, in this paper, we propose an attention based method which can learn to find spatial regions containing more useful information for detecting PA and meanwhile suppress less useful ones. In order to further improve the performance, we introduce exclusivity regularization to reduce the redundancy between different attention maps, and employ ranking loss to better fuse the classification results on the obtained multiple attention maps. The proposed network can be trained effectively in an end-to-end manner. Intra-evaluation experiments on Oulu-NPU dataset and cross-testing experiments between CASIA-MFSD and Replay-Attack show that the proposed method achieves competitive results compared with the state-of-the-art.