Hierarchical Federated Learning in Inland Waterways via Dataset Distillation and Resource Allocation

The goal of 6G is to achieve ubiquitous connectivity in heterogeneous and large-scale networks, and Hierarchical Federated Learning (HiFL) is an effective approach to assist AI technologies at the network edge. However, the “waist-shaped client-channel-cloud” architecture of maritime communication network greatly limits the implementation of hierarchical federated edge learning. Considering the aforementioned limitations, a hierarchical federated data distillation method based on UAV-assisted MEC systems is proposed, called Data Augmentation Distilled Federated Learning (DADFL). Specifically, DADFL reduce the communication process between clients and edge servers to just one round. Each client extracts their private dataset, sends synthesized data to the server, and collectively trains a global model. Building upon this, we propose a Global Optimal Iterative Search Algorithm (GOISA) considering client heterogeneity and resource constraints. GOISA minimizes overall latency by optimizing client heterogeneity, bandwidth allocation, and UAV transmission rates. Simulation experiments demonstrate that our proposed system achieves performance comparable to or even better than traditional federated learning, while reducing communication overhead by three orders of magnitude, greatly improving communication efficiency.