Attack Detection and Identification Method for Autonomous Vehicle Localization Systems Based on Gaussian Processes

The localization system plays a critical role in unmanned vehicles but is vulnerable to sensor attacks that pose a serious risk of accidents. Therefore, it is essential to detect and identify such attacks to ensure the security of unmanned vehicles. However, existing approaches fail to adapt to the dynamic driving scenarios in autonomous vehicles because they rely on fixed anomaly boundaries limited to specific scenarios. To overcome these challenges, we propose a novel real-time adaptive attack detection, identification and data recovery method in unmanned vehicles. The vehicle dynamics model is first combined with a Gaussian process (GP) to establish a probabilistic model for the localization system of autonomous vehicle. Secondly, linear regression is employed to optimize further the anomaly boundary predicted by the GP. Finally, the measurement sensor data is compared with the anomaly boundaries to obtain an anomaly report, which achieves attack detection and localization and replaces the measurements with the predicted data of GP. Compared with previous methods, the proposed method can dynamically adapt to various autonomous driving scenarios in real-time and requires a smaller number of unlabeled training samples. We evaluate the effectiveness and adaptivity of the proposed method against various sensor attacks using real-world unmanned vehicle data in complex driving scenarios. Compared to the previous method, our method improved accuracy by an average of 7.12%, reduced false negative rate by 3.29% and reduced time consumption by 33.16%.