BAsyncFed-IDS: A buffered asynchronous federated learning-based intrusion detection system for in-vehicle networks

The Controller Area Network (CAN) is highly susceptible to malicious attacks due to its lack of encryption and authentication mechanisms, leading to the necessity of intrusion detection systems (IDS). However, centralized data processing not only incurs high storage costs and bandwidth overhead but also aggravates privacy risks. This context makes federated learning (FL) a more feasible alternative. Most existing FL-based IDSs rely on synchronous aggregation methods, which overlook real-world factors such as the significant disparity in computing resources among vehicle clients, unstable vehicle network communication, and non-independent and identically distributed (non-IID) data samples. This oversight results in reduced model training speed and detection accuracy. To address the limitations of existing methods, this paper proposes BAsyncFed-IDS, a novel buffered asynchronous federated learning-based intrusion detection system. Specifically, the framework stores completed local updates in a buffer, selects models according to weight divergence, and uses mixing weights determined by staleness to reduce the influence of delayed updates. For local detection, we employ Lite-MobileNet, a lightweight detector suitable for resource-constrained in-vehicle devices. A unified image representation is used to standardize CAN traffic from different vehicles, thereby supporting intrusion detection across multiple vehicle models. Comprehensive experiments on three public datasets demonstrate that BAsyncFed-IDS can train a global model under heterogeneous vehicle data and delayed client updates, achieving favorable convergence and detection performance. It outperforms various baseline methods in convergence speed and generalization ability, thus providing a novel, accurate, and efficient intrusion detection solution for in-vehicle network security.