Context-Aware Face Presentation Attack Detection (A Dual-Branch Convolutional Neural Network)
Author:
Tariq Ahmed Bahatheq
Independent AI Researcher, Saudi Arabia
doi.org/10.52132/Ajrsp.e.2025.78.1
Face recognition systems are susceptible to presentation attacks, which can severely compromise their reliability in security-sensitive applications. Existing methods, such as Deep Pixel-wise Binary Supervision (DeepPixBis), primarily focus on facial regions, often neglecting critical contextual cues in the surrounding image that could signal spoofing attempts. This paper introduces an efficient dual-branch convolutional neural network architecture that integrates facial and contextual information for robust face presentation attack detection, all while maintaining a compact model size. The proposed model processes the extracted face and the entire image independently, producing a pixel-wise map for the face and a binary output for the full image. Trained and evaluated on the OULU-NPU dataset using standard ISO/IEC 30107-3 metrics, the proposed approach achieves state-of-the-art performance among DeepPixBis-based models in protocols II and III. Additionally, it demonstrates state-of-the-art performance in protocol II across all existing models, not just those based on DeepPixBis. Remarkably, it achieves this while being the smallest model among all existing anti-spoofing deep-learning models (1.4M parameters), demonstrating its practicality in real-world scenarios.
Keywords:
Anti-spoofing, Facial recognition, Presentation Attack Detection, Deep Learning, Dual-Branch Network, OULU-NPU
Academic Journal of Research and Scientific Publishing
