WIP: Adversarial Object-Evasion Attack Detection in Autonomous Driving Contexts: A Simulation-Based Investigation Using YOLO

Engineering notes

Key topics: autonomous driving, object detection, carla, perception. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

—Physical adversarial objects-evasion attacks pose a safety concern for automated driving systems (ADS) and are a significant obstacle to their widespread adoption. To enhance the ability of ADS to address such concerns, we aim to propose a human-AI collaboration framework to bring human in the loop to mitigate the attacks. In this WIP work, we investigate the performance of two object detectors in the YOLO-series (YOLOv5 and YOLOv8) against three physical adversarial object-evasion attacks across different driving contexts in the CARLA simulator. Using static images, we found that YOLOv8 generally outper-formed YOLOv5 in attack detection but remained susceptible to certain attacks in specific contexts. Moreover, the study results showed that none of the attacks had achieved a high attack success rate in dynamic tests when system-level features were considered. Nevertheless, such detection results varied across different locations for each attack. Altogether, these results suggest that perception in autonomous driving, the same as human perception in manual driving, might also be context-dependent. Moreover, our results revealed object detection failures at a braking distance anticipated by human drivers, suggesting a necessity to involve human drivers in future evaluation processes.

Links and sources

• Official / arXiv page

Comments