# Tackling Non-IIDness in HAPS-Aided Federated Learning

arXiv: https://arxiv.org/abs/2401.05308 · HTML: https://mlsanity.com/paper/2401.05308/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 5 of its 19 references that are in the map.
- First posted on arXiv 10 Jan 2024.

## Builds on (most cited first, 5 of 5)

- [Federated Learning on Non-IID Data: A Survey](https://mlsanity.com/paper/2106.06843/) (2021)
- [Decentralized Federated Learning: Fundamentals, State of the Art, Frameworks, Trends, and Challenges](https://mlsanity.com/paper/2211.08413/) (2022)
- [Stochastic Client Selection for Federated Learning with Volatile Clients](https://mlsanity.com/paper/2011.08756/) (2020)
- [Survey on Near-Space Information Networks: Channel Modeling, Transmission, and Networking Perspectives](https://mlsanity.com/paper/2310.09025/) (2023)
- [FLSTRA: Federated Learning in Stratosphere](https://mlsanity.com/paper/2302.00163/) (2023)

## Abstract (arXiv)

High-altitude platform stations (HAPS) enable large-scale federated learning (FL) in non-terrestrial networks (NTN) by providing wide-area coverage and predominantly line-of-sight (LoS) connectivity to many ground users. However, practical deployments face heterogeneous and non-independently and identically distributed (non-IID) client data, which degrades accuracy and slows convergence. We propose a weighted attribute-based client selection strategy that leverages server-side indicators: historical traffic behavior, instantaneous channel quality, computational capability, and prior-round learning contribution. At each round, the HAPS computes a composite score and selects the top clients, while adapting attribute weights online based on their correlation with validation-loss improvement. We further provide theoretical justification that traffic-derived uniformity can serve as a proxy for latent data heterogeneity, enabling selection of client subsets with reduced expected non-IIDness. Simulations demonstrate improved test accuracy, faster convergence, and lower training loss compared with random, resource-only, and single-attribute baselines.