Heterogeneous Architectures for Machine Learning–Based Intrusion Detection in Autonomous Vehicles

Connected and autonomous vehicles are emerging as a viable option to improve traffic flow, reduce bottlenecks, and enhance safety. In this context, the vehicle's computer system must be efficient in both performance and energy use, e.g., by identifying objects, changing speed, or operating during braking. Onboard computer architectures are heterogeneous, e.g., CPUs, GPUs, and FPGAs, and are interconnected with sensors for collection, processing, and communication. Communication flows through edge, fog, and cloud systems that comprise a smart city and telemetry system. However, communication can create gaps that intruders can exploit to attack the system and compromise its security and proper functioning. In this way, the problem addressed centers on the potential cybersecurity flaws in autonomous vehicles, which need to be identified and characterized to ensure a secure computing system. The hypothesis is that analyzing and characterizing threats to the vehicle enables effective monitoring and yields sufficient data for an effective computer architecture design for intrusion detection, without degrading computational performance or energy consumption. Therefore, a data-driven project converges to an application-specific architectural contribution based on machine learning. The project's methodological strategy establishes a systematic, in-depth review of threat scenarios and cyber-attacks on autonomous vehicles to define experiments for data collection, heterogeneous architecture design, and performance, energy, and accuracy evaluation. The materials include the following: the CARLA simulator, widely used for research into autonomous vehicles; the SIMCenter PUC Minas, which has a computational system integrated into the cockpit of a Renegade, capable of simulating scenarios and generating real data; and well-known public datasets, CPUs, GPUs, and FPGAs. This is a CNPq-funded cooperation project hosted at PUC Minas, involving other Brazilian universities (UNIFEI, UFS, and UFPA) and UC Irvine (USA). This project is currently under development and welcomes contributions from researchers and institutions.