From On-Board Data to Network Knowledge: A Group-Based Federated Learning Approach for Vehicular Networks

While intelligent transportation systems rely heavily on real-time data from vehicles, traditional centralized learning approaches face challenges in terms of privacy concerns and communication overhead in large vehicular networks. Federated learning (FL) offers a promising alternative by enabling collaborative learning on distributed data. In this paper, we propose a novel group-based FL approach for vehicular networks, where vehicles are first grouped based on real-time geographical location. Within each group, a few rounds of FL are conducted using on-board data to generate a representative model. A cluster head (CH) is then selected within the group based on a quality metric for the local models. As vehicles in proximity share similar sensed data and likely face similar decisions, only the CH, on behalf of the group, will be involved in future training and communication rounds. This reduces the communication and computation resources required for all vehicles to participate. Additionally, edge nodes, such as roadside units (RSUs), can act as intermediate aggregators between the CHs and the cloud, which significantly reduces communication latency. To further optimize this process, matching groups to edge nodes with sufficient capacity to handle the expected workload is formulated as a bipartite matching problem. Finally, the RSU-aggregated models are sent to a central server for further aggregation and global model updates.