FEDSA-ResnetV2: An Efficient Intrusion Detection System for Vehicle Road Cooperation Based on Federated Learning

Federated learning (FL)-based intrusion detection systems (IDSs) for vehicle road cooperation have attracted significant attention in recent years. However, the non-independent and identically distributed (Non-IID) data across different edge devices, coupled with uneven computational power, impedes the detection accuracy and efficiency of FL-based IDSs deployed on such devices. To address these issues, this study introduces an efficient IDS to protect privacy of vehicle users using an improved residual network and a novel FL framework, named FEDSA-ResnetV2. Firstly, the improved residual network SA-ResnetV2 is proposed, which adopts a core submodule named SA-ResnetV2-depthwise (SRD) that integrates the advantages of MobileNetV2, ShuffleNet V2, and Self-Attention mechanism for capturing key features of traffic data to effectively detect attacks. Secondly, the novel FL framework is built, which selects the optimal model globally to enhance detection accuracy, and integrates 8-bit post-quantization technique for resource-efficient edge deployment. To further improve FL training efficiency, an early stopping mechanism is employed. Experimental results show that FEDSA-ResnetV2 outperforms 11 recent Non-FL baselines and 3 state-of-the-art FL baselines in detection performance on three datasets, achieving a maximum performance increase of up to 34%. In terms of execution efficiency, it surpasses the optimal baseline in inference time, model size, mFlops, and parameters, reducing them by up to 72.6 times.