Clustered federated learning (CFL) is a promising solution to address the performance degradation posed by inconsistent data distribution, enhancing model collaborative training through homogeneous data clustering. However, the dynamic nature of user mobility in mobile edge computing (MEC) scenarios results in highly inconsistent data distributions, further complicating collaborative training. So far, existing CFL schemes rely on a central server for clustering, while the prominent clustering algorithms require a predetermined number of clusters. This inevitably leads to a relatively fixed cluster structure, which fails to adapt to user mobility, thereby hurting the model's performance. To capture the inconsistent data distribution in the dynamic user mobility scenario, this paper proposes a decentralized clustered federated learning framework (FedCVM) based on the voter model, which comprises two modules: an iterative adaptive clustering algorithm (ACAS) and a transition probability-based neighbor aggregation algorithm (TPNA). Specifically, ACAS is designed based on the greedy strategy of minimum spanning forest, which achieves optimal cluster division without predetermining the number of clusters, to realize user dynamic clustering. To further improve the model accuracy, the proposed TPNA sets the optimal exchange weights based on the neighbor interaction principle of the voter model to implement precise updates of similar user parameters. We have conducted the theoretical analysis to demonstrate that FedCVM can select the optimal cluster head and ensure model convergence. Extensive experiments on MNIST, CIFAR-10, DermaMNIST, and HARBox demonstrate that FedCVM consistently achieves the highest accuracy while maintaining a favorable trade-off between accuracy and time cost. Moreover, under simulated mobile scenarios, it yields at least a 5.1% improvement in accuracy and achieves at least a 1.29x reduction in communication overhead compared to SOTA methods. We further validate FedCVM through real-world deployment on a heterogeneous edge testbed, demonstrating its effectiveness in dynamic, decentralized mobile scenarios.

Federated learning (FL) provides a promising solution to meet the requirements of data privacy and security in intelligent transportation systems (ITS), which enables edge devices and road side units (RSUs) to collaboratively train learning models without exposing the raw data. However, the deep leakage from gradients (DLG) still leads to the risk of divulging the original data. Meanwhile, the existing gradient protection methods based on secure multi-party computation (SMC) result in huge communication overheads and latency, which are difficult to satisfy the real-time demands of both FL and the diverse services in ITS. Focusing on improving edge data security in ITS, this paper proposes an enhanced federated learning (FL) model (SemBroc-RF) with reinforcement learning, which considers the advantages of both end-to-end homomorphic encryption (HE) and SMC. To reduce communication overheads and strengthen data security in RSUs and end devices simultaneously, a partially encrypted secure multi-party broadcast computation algorithm (SemBroc) is designed, which achieves the time complexity O(n) by constructing the decoding function and sharing the gradients among the local models. To improve the model accuracy by gradients aggregation, a FL algorithm (GreFLa) with reinforcement learning is proposed based on the adaptive assigned weight of the local gradients. Theoretical analysis and detailed simulation results verify that SemBroc-RF can effectively prevent gradient leakage. On the MNIST and CIFAR-10 datasets, compared with the benchmark, the accuracy of SemBroc-RF is increased by 3.63% and 1.35%, and the training round of SemBroc-RF is reduced by 70.8% and 45.6%, respectively.