Towards Efficient Learning Using Double-Layered Federation Based on Traffic Density for Internet of Vehicles

It is an important topic to research on the federal-learning based smart services to achieve data privacy preservation in the Internet of Vehicles field. However, model training in the vehicles is still confronting the challenge of low learning efficiency when applying the federal-learning concept into the scenario of dense road. To address the above issue, this paper presents a novel technique to enhance the learning efficiency based on traffic density for the Internet of vehicles. First, a double-layered federation architecture is built through coordinating multiple roadside units. The streams of traffic are divided into different regions, where the devices inside each region are federated for down-layer learning. The roadside units corresponding to each region layer are federated for up-layer learning. Second, based on the double-layered federation architecture, an efficient federal-learning algorithm is invented, where the computational overheads of dense traffic are decreased and the data privacy is still preserved during the model training process. Finally, the simulations are conducted using the real-world dataset from the Microscopic vehicular mobility trace of Europarc roundabout, Creteil, France. The simulation results show that the proposed efficient federal-learning algorithm can improve the learning performance and preserve data privacy in the scenario of intensive traffic.