ARMBench: Benchmarking Adversarial Robustness of Multitask Perception in Autonomous Driving

Abstract

Perception models are fundamental to autonomous vehicles, enabling critical tasks such as object detection, drivable area segmentation, and lane line segmentation. Recently, multitask perception models that unify multiple vision tasks into a single architecture have shown their promise in efficiency and improved performance. However, their robustness under adversarial conditions remains underexplored. In this work, we present ARMBench, a benchmark designed to evaluate the adversarial robustness of multitask perception models for autonomous driving. ARMBench integrates a suite of representative attack methods as well as multiple defense mechanisms. Through a case study on the YOLOP model using the BDD100K dataset, we investigate the impact of adversarial attacks on individual tasks, compare the robustness of multitask and single-task models, and assess the effectiveness of defense strategies. Our results reveal asymmetric vulnerabilities across tasks and highlight the need for task-aware evaluation and defense. We have open-sourced ARMBench at https://github.com/Yunge6666/YOLOP-adversarial-attack-and-defense to contribute to the research community by providing a reproducible, extensible foundation for advancing robust and trustworthy multitask perception in AV systems.