Introducing Intelligent Data Sharing to Vehicular Cooperative Federated Learning

This paper proposes a simple yet unexplored measurement and federated learning system architecture for connected vehicles. The novelty of the introduced system is to combine the real-time data-sharing of crowdsensing with federated learning of global traffic models, providing up-to-date information for decision-making, facilitating faster learning, improving communicational channel usage, and possibly enhancing data privacy. This multi-level cooperative federated learning system generally supports operational, tactical, and strategic planning; therefore, we demonstrate its merits with a case study of parking monitoring in a simulated town as well as average speed prediction in a simulated part of Hannover, Germany. However, real-time data-sharing is essential for decision-making; it might also contain privacy-sensitive information regarding the trajectory of the vehicles. To mitigate the risk of privacy leakage, we experimented with different data selection methods for data exchange, introducing an optimization method inspired by Zeuthen’s negotiation strategy. We also checked the privacy impact of real-time data-sharing on federated learning. Our results indicate only negligible differences in privacy leakage between the proposed data selection methods. On the other hand, real-time data-sharing improves the reaction time of the federated learning system. The Zeuthen-inspired optimization method can efficiently supply valuable information for the communication partners. Moreover, it can enhance privacy protection in federated learning in some cases.