Face anti-spoofing is an essential component of modern biometric authentication systems, yet existing models often exhibit performance degradation for underrepresented ethnic groups. This thesis investigates the presence of ethnic bias in face presentation attack detection (PAD) and proposes a lightweight, real-time solution using the MiniFASNet architecture. A new dataset, SARSpoof, consisting of spoof and bona fide samples from South Asian individuals, was developed to address demographic imbalance in conventional training data. Extensive experiments demonstrate that MiniFASNet achieves over 90% accuracy on replay and print attacks, while significantly reducing bias for South Asian faces. The work also includes a fully implemented GUI-based attendance system integrating PAD and face recognition modules. The findings highlight the importance of fairness-aware design in biometric security and offer a practical framework for improving inclusiveness in real-world face authentication systems.
Raoha Bin Mejba (Fri,) studied this question.