Evaluating the Efficacy of Adversarial Recovery Methods in Traffic Sign Classification for Autonomous Vehicles
Doreen Sebastian Sarwatt, Frank Kulwa, Adamu Gaston Philipo, Jianguo Ding, Xuanxia Yao, Huansheng Ning · 2025
In Autonomous Vehicle (AV) technology, the robustness of Traffic Sign Classification (TSC) systems is crucial for ensuring safe navigation. Currently, TSC systems lack dedicated adversarial recovery methods, making them susceptible to adversarial attacks. This study embarks on an innovative path by adapting six established adversarial recovery methods from general image classification (IC) to TSC in AVs. Prompted by the limited availability of TSC-specific adversarial recovery solutions, our research undertakes a comparative analysis of these six methods to evaluate their applicability and effectiveness in the TSC domain. Each method is meticulously adapted and integrated, considering the unique challenges and requirements of TSC in AV environments. Our findings reveal insights into their performance, with the Purifying Variational Autoencoder (PuVAE) method outperforming others, achieving recovery rates of 96.15%, 79.24%, and 60.18% for Chinese, Belgium, and Germany traffic sign datasets, respectively. Additionally, it demonstrated Structural Similarity Index Measure (SSIM) values of 0.58, 0.53, and 0.66, along with recovery times of 0.0009 seconds, 0.0008 seconds, and 0.0008 seconds for the respective datasets. By highlighting effective recovery methods and addressing the unique challenges of adversarial attacks in TSC, this study contributes to improving the safety and resilience of autonomous driving systems. This research not only fills a significant gap in the safety mechanisms of AVs but also paves the way for future exploration and development of more robust and secure TSC systems, capable of effectively countering a wide range of adversarial attacks.