Hybrid AI Models for Threat Prediction and Mitigation in Autonomous Vehicle Networks

Abstract

All of the above threats - cyber and traditional - manifest as systemic safety of life risks before, during, and after the event. They can also apply during all modes of a CAV operation, from manual driving up to fully autonomous driving, making it important for future CAV fleets to exhibit resilient operation and high resistance to these aforementioned threats. Adding to these constraints, the CAV's decision-making ability is expected to react in real time, embody high trust levels, and limit all possible safety risks, all of which are particularly hard obligations to satisfy when dealing with large fleets of vehicles. While multiple trials and demonstrations showed that CAV system shortcomings can be solved by refining the fleet operation logic and artificially increasing the human intervention percentage, also a series of AI-driven and human-in-the-loop methodologies have been introduced and studied that aim to mitigate the specific weaknesses for four traditionally difficult threat groups that largely apply in the context of CAV fleets: cyber threats, passive safety threats, active safety threats, and poisoning threats. In this chapter, we use the term passive safety threat to refer to uncommon events that cause the target system to either suffer system degradation or to fail catastrophically when reactively aggregated with common operational conditions and characteristics.