WIP: Understanding the Mechanisms Behind NDT-Based Localization Vulnerabilities in Autonomous Driving

Accurate localization is critical for autonomous driving (AD), yet its security risks remain insufficiently explored, particularly in driving scenarios involving sensor fusion. This study investigates the vulnerabilities of Normal Distributions Transform (NDT) scan matching, a widely used localization method, and analyzes the conditions under which localization errors occur. We reveal that NDT relies primarily on nearby LiDAR point cloud structures from the pre-built map, making it susceptible to gradual manipulations that accumulate over time. To evaluate the impact of such manipulations, we conduct experiments simulating real-world scenarios, incorporating sensor fusion with an Extended Kalman Filter (EKF). Our findings identify key factors influencing localization errors, including target object selection and movement patterns, and confirm that these manipulations can induce errors of up to 23 m. End-to-end evaluation demonstrates that these errors can lead to hazardous driving behaviors, such as lane departures, missed traffic signals, and unintended sidewalk encroachments. By systematically analyzing the vulnerability of NDT-based localization, this study highlights the need for more robust localization mechanisms in AD.