Fed-CALiBER: Federated Lightweight BERT Intrusion Detection on CAN Bus Protocol in Autonomous Vehicle

Abstract

Recent advancements in autonomous vehicle (AV) technology have highlighted critical cybersecurity vulnerabilities within In-Vehicle Networks (IVNs), particularly the Controller Area Network (CAN) bus. While numerous Intrusion Detection Systems (IDS) exist, significant gaps remain in addressing resource efficiency and the challenge of Non Independent and Identically Distributed (Non-IID) data in distributed vehicular environments. This study proposes Fed-CALiBER, a novel framework that synergistically combines a compact, pre-trained Lightweight BERT model with a Federated Learning (FL) architecture. By training collaboratively on distinct datasets assigned to Raspberry Pi edge clients, our approach preserves data privacy by keeping raw data localized and is explicitly designed to enhance generalization across Non-IID data distributions. With a reduced communication overhead by transmitting a small parameter footprint (approx. 13 MB) during federated updates, Fed-CALiBER minimizes network overhead during parameter aggregation. Our two-cycle experimental results demonstrate that the federated global model significantly outperforms standalone models in cross-dataset generalization—improving F1-scores on unseen datasets from as low as 71.39% to over 96.59%—and successfully adapts to shifting data distributions. The framework is validated as a practical edge solution, achieving real-time inference (3–4 ms per sample) with low computational overhead on Raspberry Pi clients, representing a lightweight edge client.