AutoTAG:
A Testbed to Generate Automotive Attack Datasets

Summary

Autonomous driving is promising to revolutionize personal transportation. It relies on accurate sensor data and complex AI-based decision-making pipelines to perform safety-critical maneuvers, but recent research has demonstrated that these systems are vulnerable to adversarial sensor manipulations. However, prior work uses diverse testbeds or simplified simulation environments, making results difficult to reproduce, compare, and independently verify. In addition, prior work is often only evaluated on isolated modules or targets a single sensor type. Thus the impact of such attacks on downstream driving behavior or realistic multi-sensor fusion pipelines often remains unclear. Furthermore, attack datasets that could be used for training machine learning models or developing defense approaches are largely unavailable. Hence, there is a need for a framework that enables reproducible system-level attack evaluation while automatically generating labeled datasets. In this work, we propose AutoTAG, a framework that combines Carla and Autoware, an automotive environment simulator and an autonomous driving platform, to enable automated generation of datasets, including environmental sensor data and internal processing data of the autonomous vehicle. AutoTAG enables reproducible system-level evaluation of benign and attack scenarios and provides an interface to reproduce attacks from prior work. The resulting datasets include frame-wise attack ground-truth labels to allow easy use in machine learning applications, e.g., for attack detection training. We demonstrate the capabilities of AutoTAG with several different attacks, and provide the related datasets and artifacts.