CATI – An Open-Source Framework to Evaluate Attacks on Cameras for Autonomous Vehicles

Cameras and the subsequently applied perception algorithms are essential for the safe operation of autonomous vehicles. While many attacks on this processing pipeline are known, their impact is often evaluated on generic, non-automotive Machine Learning models. Although such models are still widely used in research, a realistic attack evaluation is not possible with them. A central problem for security researchers is the lack of realistic open-source Machine Learning models that represent autonomous driving functionalities. In our work, we propose CATI, an open-source framework to evaluate attacks on cameras in autonomous vehicles. Besides two trained models for automotive object detection and traffic sign detection, it is designed to be modular to include further models for other tasks. The two integrated models are specifically trained versions of the established object detection model YOLO. We show different attacks that successfully trick commonly available implementations of YOLO but not our trained models. Additionally, we highlight how the model robustness benefits in challenging real-world scenarios.