Abstract : The protection of biometric systems from presentation attacks involving printed photos, video replays and 3D masks depends on face anti-spoofing technology. A novel proposed work based on Graph Neural Networks (GNNs), Transformer-based feature extraction alongside Reinforcement Learning (RL) enables dynamic multi-modal (RGB, depth, infrared) data fusion for advanced spoof detection. The Proposed work executes three interconnected components which include GNN for complex inter-modal relationship understanding and Transformers for global dependency detection and RL for real-time fusion strategy optimization. The proposed work shows superior performance across three popular datasets including CASIA-SURF, Replay-Attack, and OULU-NPU by achieving Half Total Error Rates (HTER) of 6.9%, 9.8% and 6.2% respectively while producing results better than existing methods by a wide margin of 3.2%. Ablation tests prove the significant contribution of GNNs to the system by revealing a 1.8% HTER increase but RL enablesthe system to function with 0.6% worse results. The proposed work demonstrates 3.9% Attack Presentation Classification Error Rate (APCER) and 3.8% Bona Fide Presentation Classification Error Rate (BPCER) on CASIA-SURF along with a 3.9% ACER which proves its ability to detect various attack types. The study demonstrates how using relational modeling with global context learning and adaptive fusion efficiently supports secure face authentication processes.

Keyword : 3D masks, Classification Error Rate, Graph Neural Networks, Reinforcement Learning, Transformer-based feature extraction.