FedCoBott: Communication-Efficient Personalized Federated Learning for UWB LoS/NLoS Classification
IEEE Open Journal of the Communications Society · 2026 · 7 · 13828-13847
DOI: 10.1109/ojcoms.2026.3740787 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Open Journal of the Communications Society, 2026. Verfasst von Phuong B. Duong und 5 weiteren Autor:innen. Die Arbeit nennt 48 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: European Union’s Horizon Europe Research and Innovation Actions.
Abstract (Originalsprache)
Ultra-wideband (UWB) localization is highly sensitive to multipath propagation and non-line of sight (NLoS) conditions, making accurate line of sight (LoS)/NLoS classification essential for robust positioning. Existing data-driven approaches commonly rely on centralized training, which requires transferring large volumes of raw channel impulse response (CIR) data and limits scalability in distributed deployments. We propose FedCoBott (Federated Communication Bottleneck), a communication-efficient personalized federated learning framework that exchanges only a compact layer while keeping the remaining model components client-specific. FedCoBott-Single processes raw CIR measurements and shares a bottleneck layer, whereas FedCoBott-Dual combines raw CIR measurements with CIR-derived features and shares a fusion layer. We evaluate FedCoBott using real-world UWB datasets under anchor-level, sub-environment-level, and cross-environment-level heterogeneity. Conventional federated baselines, including Federated Averaging (FedAvg) and Federated Proximal (FedProx), exhibit noticeable degradation under heterogeneous client data, whereas FedCoBott substantially mitigates this degradation and achieves near-centralized performance in several settings. Comparison with local-only training shows that the benefit of selective sharing is small in the nearly saturated anchor-level setting, clearer at the sub-environment level, and architecture-dependent across environments, where local-only dual-branch training remains more accurate. FedCoBott reduces communication by approximately 98.2% for the single branch and 96.1% for the dual branch relative to FedAvg and FedProx, and by up to 99.9% relative to centralized raw-data upload, while keeping raw data local. These results demonstrate an effective accuracy–communication trade-off for distributed UWB LoS/NLoS classification.
Förderung
- European Union’s Horizon Europe Research and Innovation Actions
Zitieren
APA
Duong, P. B., Navidan, H., Fontaine, J., Bröring, A., Shahid, A., & Poorter, E. D. (2026). FedCoBott: Communication-Efficient Personalized Federated Learning for UWB LoS/NLoS Classification. IEEE Open Journal of the Communications Society, 7, 13828-13847. https://doi.org/10.1109/ojcoms.2026.3740787
Vancouver
Duong PB, Navidan H, Fontaine J, Bröring A, Shahid A, Poorter ED. FedCoBott: Communication-Efficient Personalized Federated Learning for UWB LoS/NLoS Classification. IEEE Open Journal of the Communications Society. 2026;7:13828-13847. doi:10.1109/ojcoms.2026.3740787
BibTeX
@article{Duong2026,
author = {Duong, Phuong B. and Navidan, Hojjat and Fontaine, Jaron and Bröring, Arne and Shahid, Adnan and Poorter, Eli De},
title = {FedCoBott: Communication-Efficient Personalized Federated Learning for UWB LoS/NLoS Classification},
journal = {IEEE Open Journal of the Communications Society},
year = {2026},
volume = {7},
pages = {13828--13847},
doi = {10.1109/ojcoms.2026.3740787}
}
EUVIMED