Novel presentation attack detection algorithm for face recognition system: Application to 3D face mask attack

The face biometric systems are highly vulnerable for the presentation attack that can be carried out by presenting a photo or video or even a 3D mask. In this paper, we present a novel Presentation Attack Detection (PAD) algorithm that can accurately detect and mitigate the 3D mask attacks on a face recognition system. The proposed scheme extracts both local and global features from the captured face image. The local features employed in this work corresponds to the eye (periocular) and nose region that are expected to provide clue on the presence of the mask. In addition, we also capture the micro-texture variation as a global feature using Binarized Statistical Image Features (BSIF). We then train a linear Support Vector Machine (SVM) independently on these two features whose scores are fused using the weighted sum rule before making the decision about a real face or an artefact. Extensive experiments are carried out on the public 3D mask database 3DMAD that shows the superiority of the proposed scheme with an outstanding performance of HTER = 0.03%.