Lightweight Hybrid Architecture for Secure Facial Authentication Combining Neural Networks, Liveness Detection, and rPPG

• DOI: https://doi.org/10.22533/at.ed.13176526040512

• Palavras-chave: ..

• Keywords: ..

• Abstract: Face spoofing attacks remain a major vulnerability in facial authen- tication systems based only on spatial cues. This paper proposes a lightweight RGB-only architecture that combines CNN, liveness, and rPPG signals. The three modules are combined through weighted score-level fusion. Evaluation was performed in two different scenarios, with 100 samples each, comprising bona fide presentations and attacks. The proposed fusion achieved APCER = 2.0%, BPCER = 2.0%, and Accuracy = 98.0%. Compared with the CNN-only baseline, the method reduced false acceptance by 91.7%. The results show that behavioral and physiological cues effectively correct CNN errors. The system operates at 15 - 20 fps on consumer hardware without specialized sensors.