FedLiTeCAN : A Federated Lightweight Transformer for Fast and Robust CAN Bus Intrusion Detection

TL;DR

FedLiTeCAN tackles the need for fast, robust intrusion detection on CAN buses in resource-limited intelligent vehicles by introducing a compact encoder-only Transformer deployed in a Federated Learning framework. The approach delivers a 0.40 MB model with 0.104M parameters, achieving 98.5% overall accuracy and 99.996% unseen-attack detection across Car Hacking and Survival CAN datasets, while providing edge-ready inference times around 0.608 ms per sample on Jetson Nano. Key contributions include the lightweight architecture, cross-dataset generalization to unseen attacks, and privacy-preserving FL training with FedAvg, validated against SOTA baselines. The work enables real-time, privacy-aware CAN IDS suitable for diverse vehicle types and deployment environments, demonstrating substantial gains in efficiency and robustness over prior methods.

Abstract

This work implements a lightweight Transformer model for IDS in the domain of Connected and Autonomous Vehicles

Figures (6)

• Figure 1: ECU Devices Connected through CAN Bus
• Figure 2: CAN Format
• Figure 3: Overview of Car Hacking Dataset
• Figure 5: Proposed Encoder-only Transformer Architecture
• Figure 6: Proposed Transformer in FL Environment