FSL-IDS: Federated Semi-Supervised Learning Intrusion Detection System for In-Vehicle Networks

Intelligent in-vehicle networks are increasingly exposed to complex security threats. Traditional supervised deep learning methods depend heavily on extensive labeled datasets, resulting in significant manual labeling costs, while centralized training methods incur substantial communication overhead due to the transfer of raw data. To address the dual challenges of limited labeled data and communication efficiency in in-vehicle networks, this paper proposes a Federated Semi-Supervised Learning-based Intrusion Detection System (FSL-IDS). FSL-IDS integrates a Convolutional Autoencoder (CAE) with the Federated Averaging (FedAvg) algorithm to enable a distributed and efficient intrusion detection solution. In this framework, vehicle CAN traffic data are first converted into grayscale images. The CAE is utilized for unsupervised feature extraction from unlabeled data at each local client, capturing latent representations of traffic patterns. The server then aggregates model parameters from local clients using FedAvg, forming a global autoencoder model and applying INT8 weight quantization, which reduces communication overhead by 48.7%. Subsequently, a fully connected supervised neural network is constructed atop the global encoder, requiring only 20% labeled data for fine-tuning. Experiments on the Car-Hacking dataset demonstrate that FSL-IDS achieves a detection accuracy of 98.83% and an F1-score of 94.14% across various attack types. On low-performance devices and in real vehicular environments, the inference time per sample is under 10 ms and memory consumption is only 15.28 MB. This approach provides a low-label, high-accuracy, and communication-efficient distributed paradigm for intrusion detection in in-vehicle networks, effectively balancing data utilization and edge device resource constraints.