Federated Learning (FL) is a widely used distributed learning framework for training a deep-learning global model while preserving client data privacy. This research extends FL's application into the burgeoning Internet of Things (loT) realm, particularly focusing on the Internet of Vehicles (IoV). In IoV, besides traditional privacy concerns, there is a crucial need for both global model generalization and instant localized response. We adopt a combined 3-tier FL framework and propose: (a) a Federated learning with Selected-Layer- Transmission (FedSLT) algorithm that allows partial model transmission and aggregation to cope with unreliable wireless links (which gives rise to system heterogeneity); and (b) two edge initialization algorithms Edge Aggregation (EA) and Warm-up (WU) that let edge servers shoulder more responsibility to give each edge cluster a better starting point with more customized initial model parameters (which gives rise to statistical heterogeneity) when each global round begins. Empirical experiments using different amounts of non-independent and identically distributed (non-IID) data have been done to support our theoretical analysis and illustrate the advantages of FedSLT and edge initialization algorithms over vanilla FedAvg, especially in severely heterogeneous situations.
Hierarchical Federated Learning with Edge Optimization in Constrained Networks
24.06.2024
1165351 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch