Abstract

Face anti-spoofing plays a vital role in security systems including face payment systems and face recognition systems. Previous studies showed that live faces and presentation attacks have significant differences in both remote photoplethysmography (rPPG) and texture information, we propose a generalized method exploiting both rPPG and texture features for face anti-spoofing task. First, multi-scale long-term statistical spectral (MS-LTSS) features with variant granularities are designed for representation of rPPG information. Second, a contextual patch-based convolutional neural network (CP-CNN) is used for extracting global-local and multi-level deep texture features simultaneously. Finally, weight summation strategy is employed for decision level fusion, which helps to generalize the method for not only print attack and replay attack but also mask attack. Comprehensive experiments were conducted on five databases, namely 3DMAD, HKBU-Mars VI, MSU-MFSD, CASIA-FASD, and OULU-NPU, to show the superior results of the proposed method compared with state-of-the-art methods.