Efficient rPPG-Based Biometric Authentication Using Lightweight Vision Transformer

Biometric features such as face representation and heartbeat signals from ECG are widely used for authentication. However, the existing biometric authentication systems still have defects, face recognition is susceptible to attacks such as replaying and generating videos, while ECG signals are contact detected, limiting their applications. In this study, a face video authentication technique based on rPPG is proposed. Firstly, we analyzed the facial distribution specificity of rPPG and confirmed its feasibility for identity authentication. After that, we proposed preprocessing the video data into an MSTMap, which contains both heart rate information and its facial spatiotemporal distribution, for classification. Finally, we used a lightweight Vision Transformer (ViT) for feature extraction, and to address the issue of limited training data, we employed Masked Autoencoders (MAE) for pretraining. The final algorithm achieves 97.64% accuracy, 4.33% FAR, and 4.34% FRR.