A Data-Driven Misbehavior Detection System for Connected Autonomous Vehicles

Abstract

In a Connected and Autonomous Vehicle (CAV) system, some malicious CAVs may send out false information in vehicle-to-vehicle communication to gain benefits or cause safety-related accidents. Previous false data detection methods are not sufficient to meet the accuracy and real-time requirements. In this paper, we propose a data-driven misbehavior detection system (MDS) (running by each CAV) that checks the consistency between the estimated and actually reported driving state (i.e., velocity, acceleration, brake status, steering angle) of an alerting CAV. First, MDS predicts the driving state using Gaussian mixture model based Mixture Density Network incorporating Recurrent Neural Network that can catch the driving behavior patterns of a CAV. Second, MDS extends the existing Krauss traffic flow model and uses it to consider the overall traffic flow of the road to make the predicted driving state more accurate. Finally, for a given received alert, a CAV validates the alert by checking the consistency between the predicted and actually reported driving states of the alerting CAV. We conduct extensive simulation studies based on a real driving dataset we collected from 29 participants and the Simulator for Urban MObility (SUMO) traffic simulator. The experimental results show that the false information detection rate of the proposed MDS is higher than other existing systems in different alert scenarios.