Adversarial Conditions for Autonomous Vehicle Sign Recognition: A Dataset with Original and Distorted Images
Atharv Ashish Talathi, Ayush Jain · Science Management Design Journal. · 2024
Autonomous vehicles rely on accurate sign recognition systems to navigate and respond effectively to various road conditions. However, these systems can be vulnerable to adversarial attacks, where small, often imperceptible changes to images cause significant misinterpretations by algorithms. To address this vulnerability, we introduce a dataset specifically designed to evaluate and improve the resilience of sign recognition models under adversarial conditions. This dataset includes high-quality images of standard traffic signs and corresponding distorted versions created using custom noise-generating algorithms that simulate real-world perturbations, such as added noise, visual occlusions, and other subtle alterations. The images were sourced from reputable online datasets and were resized to a uniform dimension of 32 × 32 pixels to maintain consistency for training and evaluation purposes. The dataset is structured to provide researchers and developers with a vital tool for training and testing autonomous vehicle recognition systems. By analyzing performance across both the original and altered images, researchers can enhance the robustness and reliability of these systems, making them better equipped to handle adversarial scenarios.