PSO-Based Black-Box Lane Detection Adversarial Attack

Abstract

Autonomous driving has achieved rapid development and promising performance by employing machine learning algorithms to automatically execute various maneuvers. Lane detection is one of the foremost tasks for vehicles, with a direct impact on driving decisions. Lane detection is a crucial component of autonomous driving, working in conjunction with lane centering control (LCC) and adaptive cruise control (ACC) to enable advanced driving assistance functions. At first thought, lane detection may seem immune to attacks. However, is this really the case? In this paper, we conduct experiments on lane detection modules and reveal their susceptibility to attacks exploiting hypersensitivity. Specifically, an excessively sensitive lane detector may mistake small markings for valid lanes, causing the vehicle to follow the wrong path. Through experiments, we demonstrate the vulnerability of lane detection models to adversarial attacks, even under black-box conditions where a change in one pixel can induce errors. We discuss the contrast between gradient-based and heuristic search-based optimization methods for black-box attacks and demonstrate the superiority of heuristic approaches for this task. Leveraging particle swarm optimization (PSO), we carry out lane detection attacks to address the research gap for black-box lane attacks.