Post-Processing Strategies for Detecting and Mitigating Adversarial Perturbations in Segmentation Networks
Tarek Ali, Amna Eleyan, Mohammed Al-Khalidi, Tarek Bejaoui · 2025
The safety and reliability of semantic segmentation networks face a major threat from adversarial perturbations which create problems in autonomous driving and medical imaging systems. The research develops a strong model-agnostic system which identifies and counteracts segmentation model attacks. The method applies uncertainty-based post-processing methods through pixel-wise entropy and dispersion metrics to detect adversarial inputs across different models without changing their internal structure. The method of adversarial robust knowledge distillation enables the transfer of defense capabilities from a high-capacity teacher network to an efficient student model which maintains segmentation accuracy and resource-limited deployment resilience. The integrated pipeline achieves state-of-the-art detection accuracy ($84.20 \%$) and segmentation robustness under various attack scenarios through empirical evaluation on standard benchmarks using convolutional and transformerbased segmentation architectures which outperform conventional baselines. The results demonstrate that robust distillation when used with uncertainty analysis leads to the development of reliable semantic segmentation networks for safety applications.