A Model Interpretability-Based Defense Against Clean-Label Backdoor Attacks

The widespread application of deep learning in autonomous driving and intelligent transportation has brought attention to the security of traffic sign recognition systems. Clean-label backdoor attacks present a significant threat to model robustness. This paper proposes a defense method that integrates model interpretability techniques, combining gradient-based and perturbation-based approaches. The method generates image heat maps and filters perturbed images to construct an effective data cleaning framework for identifying and removing backdoor triggers. Experiments on the Traffic Sign Recognition Dataset show that this integrated approach maintains high detection accuracy across various attack modes, achieving a Matthews correlation coefficient of 0.975 under random triggers. The method also demonstrates strong adaptability to different attack patterns. This research enhances the security of deep learning models in safety-critical applications and has broad implications for future work in securing AI systems.