Presentation Attack Detection for Face in Mobile Phones

Yaojie Liu, Joel Stehouwer, Amin Jourabloo, Yousef Atoum, Xiaoming Liu

Abstract: Face is the most accessible biometric modality which can be used for identity verification in mobile phone applications, and it is vulnerable to many different presentation attacks, such as using a printed face/digital screen face to access the mobile phone. Presentation attack detection is a very critical step before feeding the face image to face recognition systems. In this chapter, we introduce a novel two-stream CNN-based approach for the presentation attack detection, by extracting the patch-based features and holistic depth maps from the face images. We also introduce a two-stream CNN v2 with model optimization, compression and a strategy of continuous updating. The CNN v2 shows great performances of both generalization and efficiency. Extensive experiments are conducted on the challenging databases (CASIA-FASD, MSU-USSA, replay attack, OULU-NPU, and SiW), with comparison to the state of the art.