PRIVACY-PRESERVING FEDERATED LEARNING WITH LOCAL DIFFERENTIAL PRIVACY FOR TRAFFIC PREDICTION IN IOV SYSTEMS

Abstract

The Internet of Vehicles (IoV) poses severe challenges for distributed collaborative traffic prediction due to privacy concerns and scalability in the distributed learning scenario. Traditional federated learning mechanisms in vehicle networks fail to ensure data privacy while ensuring efficient consensus schemes for model aggregation. This paper proposes a novel privacy-preserving federated learning framework that integrates an enhanced Practical Byzantine Fault Tolerance (PBFT) consensus protocol with Local Differential Privacy (LDP) for trustworthy traffic prediction in IoV systems. Our new PBFT algorithm incorporates logarithmic scaling in which the consensus rate decreases as the network increases, significantly improving scalability for large vehicular networks. The design features a CNN-GRU neural network model that is trained on real PeMS traffic data, enabling accurate prediction of traffic flow, occupancy, and speed patterns. Local Differential Privacy with calibrated noise injection provides a robust privacy guarantee at 50% privacy leakage loss with an accuracy of 70-82% compared to 75-87% accuracy in non-protected models. Experimental results provide improved performance with 204ms average consensus time per hop across 200 vehicles, outperforming traditional PBFT (365ms) and alternative baselines. A multi-component incentive framework controlling accuracy, privacy, and contributions ensures equitable participation and ongoing cooperation. Our approach effectively authenticates on real California highway data with 92-95% of the centralised system's accuracy at preserving vehicle privacy, which is appropriate for realistic IoV deployments that require both efficiency and privacy preservation.