Face Presentation Attack Detection Using Remote Photoplethysmography Transformer Model

Abstract

Face Presentation Attack Detection (PAD) is essential for face recognition systems to achieve reliable verification in secured authentication applications. The face Presentation Attack Instruments include the printed photo, electronic display, wrap-photo and custom 3D masks. With the evolving technologies to generate the novel face PAI the generalisable PAD is of paramount importance. In this paper, we proposed a novel face PAD algorithm to achieve reliable detection of presentation attacks by quantifying the liveness using the remote photoplethysmography (rPPG) signal. The proposed method is developed by augmenting the PhysFormer model with an additional Temporal Difference Multi-Head Self-attention (TD-MHSA) block to obtain the reliable rPPG signal. We also proposed a novel classifier using 3DCNN to effectively capture the spatio-temporal to achieve a reliable PAD across different un-seen PAI. Extensive experiments are conducted on the publicly available OULU-NPU dataset comprised of four different PAI and six different smartphones. The proposed method is benchmarked with nine different existing PAD techniques on two different evaluation protocols and indicates considerable performance compared with the existing PAD techniques.