Abstract:
The rapid advancement of generative adversarial networks (GANs) in synthetic face image generation has made GAN-based face recognition systems vulnerable to adversarial attacks, particularly those targeting critical facial regions. This research proposes a robust detection framework to identify and mitigate these attacks, ensuring the reliability and trustworthiness of GAN-based face authentication systems. The proposed system employs a multi-component architecture, integrating computer vision and deep learning techniques. It starts with facial landmark localization, then deep feature extraction is performed to capture discriminative representations from both low-level and high-level visual cues. These features are then analyzed through multiple detection models, including adversarial key point detection (AKPD), GAN-based anomaly detection, and feature consistency checks. Advanced training strategies are employed to enhance the robustness and generalization capabilities of the detection models. Extensive experimental evaluations are conducted using multiple datasets, including real and GAN-generated face images subjected to various adversarial key region attacks. Comprehensive performance analyses are conducted using metrics like accuracy, precision, recall, F1-score, false positive/negative rates, and detection error rates. The system incorporates model interpretability techniques to provide explanations and visualizations of critical regions responsible for attack detection. The proposed adversarial key region attack detection system demonstrates superior performance in identifying and mitigating strategic perturbations targeting critical facial regions, contributing to the integrity and trustworthiness of GAN-based face recognition systems.
