Cross-Task Physical Adversarial Attack Against Lane Detection System Based on LED Illumination Modulation

Junbin Fang, Zewei Yang, Siyuan Dai, You Jiang, Canjian Jiang, Zoe Lin Jiang, Chuanyi Liu, Siu-Ming Yiu

Abstract: Lane detection is one of the fundamental technologies for autonomous driving, but it faces many security threats from adversarial attacks. Existing adversarial attacks against lane detection often simplify it as a certain type of computer vision task and ignore its cross-task characteristic, resulting in weak transferability, poor stealthiness, and low executability. This paper proposes a cross-task physical adversarial attack scheme based on LED illumination modulation (AdvLIM) after analyzing the decision behavior coexisting in lane detection models. We generate imperceptible brightness flicker through fast intensity modulation of LED illumination and utilize the rolling shutter effect of the CMOS image sensor to inject brightness information perturbations into the captured scene image. According to different modulation parameters, this paper proposes two attack strategies: scatter attack and dense attack. Cross-task attack experiments are conducted on the Tusimple dataset on five lane detection models of three different task types, including segmentation-based models (SCNN, RESA), point detection-based model (LaneATT), and curve-based models (LSTR, BézierLaneNet). The experimental results show that AdvLIM causes a detection accuracy drop of 77.93%, 75.86%, 51.51%, 89.41%, and 73.94% for the above models respectively. Experiments are also conducted in the physical world, and the results indicate that AdvLIM can pose a real and significant security threat to diverse lane detection systems in real-world scenarios.