Enhancing Adversarial Robustness in Rail Detection via Frequency Domain Denoising and Model Distillation

In view of the security risks caused by the track detection model's susceptibility to adversarial attacks and the lack of attack and defense verification on actual track defect data, this paper proposes two defense strategies: frequency domain denoising and model distillation. Frequency domain denoising combines wavelet transform and adversarial training to suppress high-frequency noise interference in adversarial samples through frequency domain decomposition. The proposed defense mechanism led to a 62.1% improvement in model accuracy and a 62.8% increase in mAP@50, demonstrating superior performance compared to using either wavelet transform or adversarial training independently. Model distillation gradually introduces adversarial samples and jointly optimizes detection loss and distillation loss. The proposed defense mechanism significantly improves model performance with gains of 13.0% in accuracy, 13.5% in recall, 20.0% in mAP@0.5, and 28.0% in mAP@0.5:0.95, establishing an optimal balance between precision and adversarial robustness.