Abstract
UAV-assisted hierarchical federated learning relies on client-UAV and UAV-BS uploads, where unreliable delivery, Non-IID data, and repeated model transmissions affect aggregation and communication use. This paper proposes Reliability-Aware Client-Vector Adaptive Dual-Level Aggregation (RCAD), integrating adaptive aggregation, bounded inverse-probability correction, cross-round weight smoothing, and probabilistic client and UAV participation. We evaluate RCAD in four controlled communication settings against FedAvg, ClientAdaptive, TwoLevelUnbiased, a hierarchical implementation of FedAWA, and Oort-H, a hierarchical participant-selection baseline, using five runs per method. Full_RCAD maintains competitive F1 scores while reducing the mean weighted attempted-upload cost by 14.8% relative to FedAvg and increasing mean F1-per-cost by 17.9%. In a scalability study from 32 clients/4 UAVs to 128 clients/16 UAVs, its normalized communication cost remains approximately 0.83–0.85. Experiments on image datasets, together with ablation and sensitivity analyses, examine the behavior of the individual components and the overall framework.
Funding Information
  • National Natural Science Foundation of China (62402450, 62672451)
  • Natural Science Foundation of Zhejiang Province (LQ24F020024)
  • 2026 University Student Innovation and Entrepreneurship Training Program (S202610345047)

This publication has 21 references indexed in Scilit: