Digital Twins Enabled On-Demand Matching for Multi-Task Federated Learning in HetVNets

In the heterogeneous vehicular networks (HetVNets), the roadside units (RUs) can exploit the massive amounts of valuable data collected by vehicles to complete federated learning tasks. However, most of the existing studies consider the scenario of one task requester (TR) and ignore the fact that multiple TRs may concurrently request their model training tasks in the HetVNets. In this paper, we consider the scenario of multi-TR and multi-RU and propose a digital twins (DT) enabled on-demand matching scheme for multi-task federated learning to address the two-way selection problem between TRs and RUs. Specifically, by jointly considering the diversified requirements of the TRs and the differentiated training capabilities of the RUs, we first design a DT enabled on-demand matching architecture to facilitate the multi-task federated learning in the HetVNets. Then, based on the personalized requirement of the DT of each TR (DT-TR), a marginal utility based vehicle selection mechanism is proposed to enable the DT of each RU (DT-RU) to determine the customized model training strategy. With the determined strategies, the two-way selection problem between the DT-TRs and the DT-RUs is formulated as an on-demand matching game in DT networks, where a matching algorithm is designed to obtain their optimal strategies. Simulation results demonstrate that the proposed scheme outperforms the conventional schemes in terms of training accuracy, performance-cost ratio (PCR), and task completion rate (TCR).