Mask Attack Detection Using Vascular-weighted Motion-robust rPPG Signals

Detecting 3D mask attacks to a face recognition system is challenging. Although genuine faces and masks show significantly different remote photoplethysmography (rPPG) signals, rPPG-based anti-spoofing methods often suffer from performance degradation due unstable alignment in the video sequence weak rPPG signals. To enhance signal motion-robust way, landmark-anchored stitching method proposed align robustly precisely at pixel-wise level by using both SIFT keypoints facial landmarks. better encode signal, weighted spatial-temporal representation proposed, which emphasizes regions with rich blood vessels. In addition, characteristics of signals color spaces are jointly utilized. improve generalization capability, lightweight EfficientNet Gated Recurrent Unit (GRU) designed extract spatial temporal features for classification. The compared state-of-the-art on five benchmark datasets under intra-dataset cross-dataset evaluations. shows significant consistent improvement over other spoofing detection.