Traffic Sign Recognition Under Visual Perturbations: Shadows, Light Patches, and Simulated Obstructions

Traffic sign recognition (TSR) systems are essential for autonomous driving yet remain vulnerable to adversarial attacks that mimic natural phenomena. These vulnerabilities pose significant safety concerns for autonomous vehicle deployment, necessitating a thorough investigation of potential attack vectors. In this paper, we investigate three distinct digital perturbation vectors: shadow manipulations, localized light patches, and simulated obstructions—all designed to maintain physical plausibility while degrading classifier performance. Unlike conventional techniques requiring physical access or producing conspicuous artifacts, our digital perturbations integrate seamlessly with the visual environment while achieving high attack success rates (ASRs). We evaluate these attacks across specialized TSR models and general-purpose CNNs on both GTSRB and PTSD datasets, enabling direct comparison between attack methodologies. Our findings reveal that CNN architectures incorporating spatial transformers provide the most vigorous defense, particularly against illumination-based perturbations. By employing targeted data augmentation and feature importance-based countermeasures, we provide actionable insights for developing more resilient TSR systems against these naturalistic adversarial attacks.