Adapting Image Classification Adversarial Detection Methods for Traffic Sign Classification in Autonomous Vehicles: A Comparative Study

Autonomous driving systems critically depend on accurately classifying traffic signs, a task that is jeopardized by adversarial attacks. This paper focuses on the relatively unexplored domain of Traffic Sign Classification (TSC) in the context of detecting adversarial attacks. We conduct a rigorous evaluation of six prominent adversarial example detection methods, each representing a distinct detection category and well-regarded within the research community. Our methodology involves adapting these techniques, originally developed for general image classification (IC), to the unique challenges posed by traffic sign images, characterized by their complexity due to factors like varying environmental conditions and a large number of classes. Our study reveals insights into their effectiveness, with the Natural Scene Statistics (NSS) method outperforming others with 83.42%, 86.42%, and 99.96% detection rates; 0.08%, 0.05%, and 0.04%, false positive rates; and 0.02sec, 0.01sec, and 0.02sec detection times for Chinese, Belgium and German traffic sign datasets, respectively. NSS’s superiority is crucial for autonomous vehicles. Our study also sheds light on the often-neglected aspect of detection time in IC, which plays a vital role in ensuring operational efficiency and safety for autonomous vehicles. Our research highlights the need for customized defense strategies tailored to the TSC domain, considering our evaluation’s findings. By identifying promising techniques for detecting adversarial attacks in TSC, we contribute to enhancing the safety and robustness of autonomous driving systems. This study fills a critical knowledge gap, providing valuable insights into understanding and defending against adversarial attacks, specifically in the TSC context.