SITAR: Evaluating the Adversarial Robustness of Traffic Light Recognition in Level-4 Autonomous Driving

Traffic Light Recognition (TLR) is vital for Autonomous Driving Systems as it supplies critical information at intersections. Modern TLRs leverage camera and geolocation data, incorporating complex pre-(post)-processing steps and multiple deep learning (DL) models for detecting, recognizing, and tracking traffic lights. While the adversarial robustness of standalone DL models has been extensively studied, the robustness of a modern TLR system, i.e., a complex software component with code and DL models, is rarely studied and hence requires research efforts.In this work, we propose a novel testing framework (namely SITAR) targeting TLR modules from a representative Level-4 ADS, such as Baidu Apollo and Autoware. We design a novel adversarial attack loss function to evaluate and improve the adversarial robustness of modern TLR systems. We applied SITAR on Apollo TLR and compared our novel loss function with the state-of-the-art approaches that can effectively attack object detection and image recognition models. SITAR is shown to be effective and our novel loss function performs better than previous SOTAs with a 93% to 100% success rate with a maximum of five-step iteration and eight pixels per perturbation.