AI-Driven Anti-Spoofing for Robust Real-Time Face Recognition

AI-Driven Anti-Spoofing for Robust Real-Time Face Recognition The security of facial recognition systems has become a critical challenge, especially in scenarios where real-time operation is required. This thesis introduces an innovative approach that integrates edge AI to prevent and counter spoofing attempts. A thorough review of the state-of-the-art reveals that many existing anti-spoofing solutions rely on additional hardware, such as costly and complex Time-of-Flight (TOF) sensors, or on traditional methods that require specific environmental conditions to function properly. These limitations hinder their scalability and practical deployment in real-world applications. To overcome these challenges, this thesis introduces an innovative edge AI approach that integrates deep learning techniques optimized for devices with limited computational resources. The proposed method employs a MobileNetV3Small backbone pre-trained on ImageNet, enhanced with additional layers to achieve accurate classification between spoof and live facial images. The system architecture is carefully designed with modules for preprocessing, feature extraction, and classification, and incorporates strategies to reduce computational complexity—such as the use of lightweight models and quantization techniques. The developed model is subsequently deployed on ST’s N6 board, which features a dedicated Neural Processing Unit (NPU) that delivers superior energy efficiency and processing speed compared to traditional CPU or GPU solutions. The training of the model is carried out on the CelebA-Spoof dataset, while validation is performed using dedicated images provided by the same dataset. The evaluation, supported by theoretical analyses and simulations, confirms the robustness and effectiveness of the proposed approach. Preliminary results, supported by theoretical analyses and simulations, highlight the potential of this approach in effectively countering spoofing attacks and enhancing the security of facial recognition systems.