A Federated LSTM Autoencoder Framework for Privacy-Preserving Intrusion Detection in V2X Networks

The fast growth of Vehicle-to-Everything (V2X) networks requires privacy-preserving Intrusion Detection Systems (IDSs) for effective operation. The proposed Federated Long Short-Term Memory Autoencoder (Fed-LSTM-AE) framework allows distributed vehicular clients to perform collaborative anomaly detection through model parameter sharing without exchanging raw data. The framework enables each client to create its own LSTM-based autoencoder model of normal traffic patterns while sharing only model parameters with a central server through federated learning to maintain data privacy and improve system scalability. Experiments using the VeReMi dataset show that Fed-LSTM-AE achieves better performance than the centralized LSTM, one-dimensional Convolutional Neural Network (1D CNN), Random Forest, and Isolation Forest baseline methods in terms of detection accuracy, F1-score, and Area Under the Receiver Operating Characteristic Curve (ROC-AUC) metrics. The framework shows strong detection performance against various attack types while achieving efficient federated training convergence and maintaining stability under non-Independent and Identically Distributed (non-IID) data conditions. The results demonstrate Fed-LSTM-AE's suitability for real-world V2X deployments because it maintains privacy protection while being adaptable and communication-efficient.