Fortifying Traffic Signal Detection Against Adversarial Attacks
Gitika Sharma, Khushi Bansal, Hansi Bhardwaj, Himanshu Sharma · 2025
Adversarial attacks pose a significant threat to the reliability and safety of machine learning (ML) models, especially in critical applications like self-driving cars. Even slight perturbations to input data can cause misclassification, undermining the trustworthiness of these systems in safetysensitive environments. This paper introduces a new adversarial dataset designed to address this issue, with a focus on adversarial training. The proposed method involves augmenting the training dataset with both clean and adversarially perturbed examples, and in some cases, performing adversarial perturbations on the entire training set. Furthermore, this work contributes to advancing traffic signal detection systems for autonomous vehicles by developing a novel pattern-matching approach that enhances model robustness. The proposed approach is aimed at minimizing false predictions and improving the accuracy of traffic signal detection in the presence of adversarial interference. This research provides insights into enhancing the resilience of self-driving car systems against adversarial attacks, contributing to the development of more reliable and secure autonomous vehicles.