Multi-server federated learning in vehicular edge computing

Abstract

Federated learning (FL) offers a promising paradigm for privacy-preserving model training in connected and autonomous vehicle networks, where vehicles act as clients and roadside units (RSUs) host FL servers at the edge. However, practical deployments face multiple challenges: highly non-IID and noisy data across vehicles, heterogeneous computation and communication resources, private training costs, intermittent connectivity, and the need to coordinate multiple servers. This thesis addresses these challenges through three complementary contributions. First, we propose T-BIDS, a dynamic, budget-aware client selection framework that explicitly accounts for client quality, cost, and availability. T-BIDS uses Thompson sampling to estimate each client’s contribution quality from validation improvements, and in each round selects participants by a quality-to-bid ratio under a per-round budget, with incentive-compatible payment rules. Clients can also augment their local datasets with targeted synthetic samples to rebalance under-represented classes. Across extensive simulations on CNNs on MNIST and GTSRB under shard- and Dirichlet-based non-IID splits with client-level label noise and mobility-induced dropouts, as well as additional experiments with ResNet-18 on GTSRB, T-BIDS converges faster and achieves higher, more stable accuracy and lower loss than baseline selectors. Second, we design an edge-based multi-server FL framework that combines performance-aware aggregation with mobility support. Each server aggregates peer updates using statistical weighting and outlier mitigation, and an application-layer handover protocol preserves model updates when vehicles move between RSU coverage areas. Compared with single-server and edge–cloud baselines, the proposed multi-server methods achieve higher accuracy and improved precision, recall, and F1-score while maintaining low latency and avoiding the additional model-transfer delays of cloud-based aggregation. Third, we propose a client-feedback-driven, quality-based reliability mechanism for model sharing in multi-server FL. In this approach, neighboring servers exchange global models, and client-side evaluations are aggregated to maintain per-peer reliability scores. These scores guide which peer model(s) are selected for reuse or dissemination as the next round initialization, enabling more effective knowledge transfer across servers—particularly in the early rounds—without requiring raw data exchange or explicit disclosure of sensitive client dataset statistics. Overall, the thesis demonstrates that coordinating clients and servers based on realtime quality, cost, and reliability yields robust, scalable, and privacy-preserving FL for vehicular edge computing.