FL4MDS: A Federated Learning Framework for Misbehavior Detection in Connected Vehicles Under Intermittent Client Participation

Abstract

In this paper, we propose a FL4MDS, a Federated Learning-based Misbehavior Detection System (MDS) for connected vehicular networks that identifies malicious and falsified messages without requiring centralized data collection. Using the VeReMi dataset, we demonstrate that a GRU-based temporal model can effectively learn sequential patterns associated with vehicular misbehavior. To address practical challenges in vehicular environments such as non-IID data distributions, intermittent connectivity, and client dropout, we extend FedProxWithSave with a checkpoint-aware federated strategy that combines the stability of FedProx with periodic model state preservation. This design enables robust training under unreliable communication conditions and allows rapid recovery from client or server interruptions. Experimental results indicate that the proposed approach maintains stable detection performance under simulated client dropout, achieving accuracy and ROC-AUC values above 96%, while providing low detection latency and high throughput. These results highlight the effectiveness and resilience of checkpoint-enabled federated learning for scalable misbehavior detection with respect to federated client participation in connected vehicle systems.