Detection of Presentation Attacks on Facial Authentication Systems Using Intel RealSense Depth Cameras

Abstract

In this paper, we explore the prospects for using Intel RealSense depth cameras to solve the problem of presentation attack detection in facial authentication systems. Studies known to date use quantized depth data. Such an approach makes it impossible to connect these data with anthropometric facial features and scene geometry. In addition, in recent papers, some researchers declared the limitation of using depth cameras in this problem only at small (less than 1 m) distances to the face. In this regard, we have made our collection of 480 samples containing bonafide images and two types of presentation attacks. The samples were taken at distances up to 2 m, as well as in different face positions relative to the camera axis. We also designed a set of features based on texture analysis, anthropometric properties, and face oval localization. The conducted experiments show the advantage of using the original, not quantized depth data. Our study also refutes the thesis that depth cameras can be used for presentation attack detection only at short distances. Additionally, we demonstrate the applicability of the proposed features on one of the publicly available datasets.