This paper presents a novel approach to address the vulnerability of traffic sign recognition (TSR) systems, which play a critical role in ensuring the safety of autonomous vehicles, to adversarial attacks. Existing defense techniques primarily focus on modifying the training process or the inference process, but remain vulnerable to attacks with a high success rate. In this paper, we propose a method to redesign traffic sign designs themselves to create signs that are both human-interpretable and resilient to adversarial attacks. To achieve this, we develop a framework that takes as input a traffic sign standard, a state-of-the-art adversarial training method, and a function that efficiently synthesizes realistic traffic sign images. It then outputs an optimized traffic sign standard for training a TSR model that is resilient to adversarial attacks. Experiments were conducted with specific implementations that modified pictograms and colors, achieving robust accuracy improvements of up to 16.33% to 24.58% compared to state-of-the-art methods. User studies confirmed that the redesigned traffic signs are easily recognizable by humans.
