Analysis of Autonomous Driving Software to Low-Level Sensor Cyber Attacks

Autonomous Vehicle (AV) architectures fuse legacy, electromechanical components with advanced sensor technology and digital controllers, governed by software. An open challenge for the design of AVs are cyber threats such as Electromagnetic Injection (EMI) attacks, to the low-level layer, comprising electromechanical components, which can propagate through to the higher-level, intelligent control, affecting decision-making and the safety of the vehicle. This study analyses the robustness of the design of the software stack of a real-world AV to attacks on the low-level actuation using the example of an EMI attack on the steering angle sensor. To achieve this, we create a hybrid testbed which combines the mathematical model of the lowlevel sensor with the high-fidelity, intelligent control. We further develop safety and performance metrics measured at the highlevel, which we use to generate a detailed view on the safety and system performance of the software. We conduct diverse EMI attacks on the target AV, within 3 diverse critical driving scenarios, consisting of $>1000$ simulations. The results indicate a correlation between an increase in attack noise with an increase in safety violations and failures to complete the mission of the AV. Our results highlight the importance for AV software developers of testing under diverse attack and driving scenarios, as each scenario within our experimentation exhibits different behaviour of the system and correlations to differing safety and system performance indicators.