Improving Machine Learning Anomaly Detection for Safety-Critical RISC-V Automotive Systems

Modern automotive systems are evolving into complex cyber-physical platforms, where traditional fixed-policy fault recovery mechanisms prove insufficient against sophisticated faults and cyber-attacks. This work presents an anomaly detection framework for RISC-V-based automotive systems, combining Hardware Performance Counters (HPC) with additional hardware metrics to improve detection accuracy under realistic conditions. The methodology is validated by running FreeRTOS workloads on a full-system RISC-V architecture with controlled fault injection using the CHAOS framework. A comparative analysis of sequence-aware and classical machine learning models demonstrates that integrating temporal data significantly enhances detection, with the GRU-Autoencoder showing the best trade-off between performance and computational efficiency for safety-critical scenarios.