Misbehavior detection system with semi-supervised federated learning

Abstract

V2X communications can enhance transportation safety by exchanging safety information between vehicles, road infrastructures, networks, and pedestrians. However, the safety messages are vulnerable to disruption from faulty components or an attack that can cause misinformation. Recently, a machine learning-based misbehavior detection system (MDS) has been widely investigated to detect the misbehaving vehicles to secure the V2X communications. Nonetheless, machine learning models need sufficient labeled data for learning purposes. However, the volume of unlabeled data is usually larger than that of labeled data in practice. Moreover, transferring the large dataset to a centralized learning model will consume much bandwidth. Thus, we propose a semi-supervised federated learning MDS to overcome the limitations of unlabeled data and bring the training close to the data sources to reduce the bandwidth to the core network. Overall, our model with only limited labeled data training (5%–30%) can achieve the F1-score up to 0.96 and the recall up to 0.95. The F1-score is up to 0.26 higher and the recall is up to 0.29 higher than the performance of centralized supervised learning. The federated learning model can reduce the core network bandwidth utilization by up to 95%.