DeepGuard-FAS: A Vision Transformer-based Face Anti-Spoofing Framework for Real-Time Biometric Security

Face anti-spoofing (FAS) is a critical component in ensuring the reliability of biometric authentication systems, especially as such technologies become integral to mobile devices, surveillance systems, and secure access control. This paper introduces DeepGuard-FAS, a robust real-time face anti-spoofing framework based on Vision Transformer (ViT) architecture. The proposed model effectively captures global contextual dependencies in facial patterns and enhances the detection of spoofing attacks such as printed photos, video replays, and 3D masks. The methodology incorporates advanced spoof pattern amplification using motion cues and local texture descriptors, along with facial alignment through MTCNN. The model was trained and tested on the MSU-MFSD dataset and validated cross-dataset using the Replay-Attack dataset. Extensive experiments demonstrate the superior performance of DeepGuard-FAS, achieving a frame-wise accuracy of 99.5% on MSU-MFSD and a cross-dataset accuracy of 99.3%, significantly outperforming existing CNN-based and ResNet models. The model also recorded an Average Classification Error Rate (ACER) of 0.005 and AUC of 0.999, showcasing its potential for deployment in real-time security applications. These results affirm the effectiveness of Vision Transformers in biometric security and open new pathways for reliable, scalable, and interpretable face liveness detection systems.