ENHANCED 3D FACE ANTI-SPOOFING FOR SECURE BIOMETRIC AUTHENTICATION USING SPATIO-TEMPORAL DEEP LEARNING NETWORKS (STDL-NET) WITH ROBUST FEATURE REPRESENTATION AND TEMPORAL CONSISTENCY MODELING

Povzetek

Face spoofing remains a significant vulnerability in facial biometric systems, where attackers employ techniques such as high-resolution photo prints, video replays, 3D masks, and even AI-generated deepfakes to deceive liveness detection modules. To counter these evolving threats, this study introduces STDL-Net, a Spatio-Temporal Deep Learning Network designed to detect facial spoofing by integrating both spatial depth cues and temporal behavioral features. STDL-Net employs a dual-stream architecture that processes RGB and 3D depth maps using 3D convolutional neural networks (3D-CNNs) to capture geometric textures and facial contours, which are inherently difficult to forge. Simultaneously, it models time-dependent facial dynamics like blinking, subtle head movements, and micro-expressions using Long Short-Term Memory (LSTM) networks, enabling detection of both static and dynamic spoofing attempts. Attention mechanisms further refine the model's focus on discriminative regions of interest, such as the eye and mouth areas. Experimental validation was conducted on a custom dataset acquired using stereo vision and structured-light sensors (e.g., Intel RealSense, Microsoft Kinect), including a balanced mix of real and spoofed samples. The proposed STDL-Net achieved high robustness, yielding 97.3% accuracy, a 0.982 AUC, and low error rates across diverse spoofing categories. These results underscore the effectiveness of combining depth sensing with temporal learning, offering a comprehensive and scalable solution for next-generation biometric security systems.