Image Analysis of Facial Blood Vessels for Anti-Spoofing of Printed Image and 3D Mask Attacks

In recent years, many face spoofing detection methods based on video frame or video sequence have been proposed to counteract the security threats of face recognition system. Although the video sequence based detection methods analyzing liveness signals have satisfactory performance, it is difficult to capture a long video in realistic applications; the video frame based methods have good real-time capabilities, but the detection performance is poor when dealing with high-resolution fake faces. Therefore, in this paper, we propose a novel video frame based face spoofing detection method exploiting facial blood flow differences, which achieves excellent detection performance and generalization capability. Specifically, compared with fake face, the subcutaneous tissue of live face is distributed with complex blood vessels. When the visible light penetrates face skin, its absorption in different regions is different due to the distribution of blood vessels. In order to extract the visible light absorption variations between different regions, the attentional face image is first fed into an 1D CNN to extract features for describing the absorption characteristics of each point; then, a GRU network is invoked to capture the absorption variations of the outputs of 1D CNN; finally, the Softmax function is used to classify whether the face image is a live face or a spoofing attack. Extensive experiments on three databases of 3DMAD, MSU and Replay-Attack demonstrate the effectiveness of our proposed method in distinguishing the printed images or 3D masks from the live ones and show that the detection performance outperforms other state-of-the-art methods.