Adversarial Object-Evasion Attack Detection in Autonomous Driving Contexts: A Simulation-Based Investigation
Research Poster Engineering 2025 Graduate ExhibitionPresentation by Rao Li
Exhibition Number 116
Abstract
Physical-world adversarial attacks present a critical safety challenge for automated driving systems (ADS), hindering their widespread deployment. To address this issue, we propose a human-AI collaboration framework aimed at achieving human-informed autonomous driving (AD). In this preliminary study, we systematically evaluated detection gaps in adversarial attack scenarios and identified key factors causing the detection failure. Specifically, we assessed the robustness of YOLOv5 and YOLOv8 against three types of physical adversarial object evasion attacks—targeting STOP signs, vehicles, and pedestrians—across diverse driving contexts in CARLA simulation platform. To obtain a comprehensive understanding of object detection distribution, we examined each attack on dynamic autonomous driving contexts at two levels: 1) the target-object level using static images and 2) the driving-context level using recorded videos. Using static images, we found that YOLOv8 generally outperformed YOLOv5 in attack detection but remained susceptible to certain attacks in specific contexts. Using recorded videos, none of the attacks achieved high attack success rates when system-level features were considered. However, the attack detection varied across locations. Altogether, the results suggest that driving context is a critical factor in autonomous driving perception and the viewpoints of targeted objects matter. Furthermore, we observed object detection failures within anticipated minimal braking distance of human drivers, underscoring the need for human involvement in future evaluations.
Importance
Our research provides critical insights into the behavior of state-of-the-art object detection models, revealing that driving context significantly influences adversarial robustness in ADS. This finding is pivotal for advancing more reliable and context-aware autonomous driving technology. Additionally, the minimal braking distance gap between human drivers’ anticipation and ADS perception highlights the necessity of integrating human-AI collaboration into the ADS development. By fostering a synergy between AI-driven perception and human factors, our work sheds light on the way for safer, more reliable automated driving systems and intelligent transportation.