PFMeta-IDS: Personalized Federated Meta-learning Automotive Intrusion Detection System with Collaboratively Adaptive and Learnable

The increasing connectivity of vehicular networks has introduced significant security challenges, particularly in safeguarding the Controller Area Network (CAN) from cyberattacks. While the CAN protocol enables efficient and low-latency data communication, its lack of built-in security mechanisms leaves it vulnerable to various attacks. Existing intrusion detection systems (IDSs) often rely on large, static datasets and centralized training, limiting their adaptability to dynamic attack scenarios and raising concerns about data privacy. To address these limitations, this work introduces PFMeta-IDS, a personalized federated meta-learning intrusion detection system. In PFMeta-IDS, the FedSWR algorithm employs similarity-weighted aggregation to balance personalization and generalization. The LDwCBN network enhances computational efficiency through the model lightweight method, ensuring suitability for resource-constrained environments. Evaluated on the Car-Hacking dataset, PFMetaIDS achieves F1-scores of 98.46% for DoS attacks, 93.68% for Fuzzy attacks, $\mathbf{9 7. 8 0 \%}$ for Gear Spoofing attacks, and $\mathbf{1 0 0. 0 0 \%}$ for RPM Spoofing attacks. These results outperform or match state-of-the-art methods. Notably, these results were achieved on local clients with low training data volumes, showcasing the system’s ability to adapt quickly while preserving data privacy. The robustness and efficiency of PFMeta-IDS make it a scalable solution for vehicular network security.