Face Anti-spoofing Based on Cooperative Pose Analysis

Face anti-spoofing has been vital to preventing face recognition systems from fake faces. However, most state-of-the-art passive methods treat face anti-spoofing as a classification problem, relying on purposive databases and well-designed backend algorithms. In this paper, we propose a novel active face anti-spoofing framework named Cooperative Pose Analysis (CPA), in which a higher cooperation degree is required in the manner of head pose changes. And we propose a new pose representation named Pose Aware Quadrilateral (PAQ), which is sensitive to pose changes of living faces and easy to identify spoof faces such as printed and twisted photographs. The proposed PAQ is easily accessed by utilizing a lightweight projective Spatial Transformer Network [11]. The whole system does not require much computational or storage resources and is easy to deploy and use. Experiments on both datasets and human subjects are conducted and indicate the effectiveness of our work.