1.
Zheng, Zijian; Deng, Yansha; Yi, Wenqiang; Shin, Hyundong; Nallanathan, Arumugam
Over-the-Air Computation Enabled Semi-Asynchronous Wireless Federated Learning Journal Article
In: IEEE Transactions on Communications, pp. 1–1, 2025, ISSN: 1558-0857.
Abstract | Links | BibTeX | Tags: aggregation optimization, Atmospheric modeling, Computational modeling, Convergence, Federated learning, Noise, Optimization, over-the-air computation, Semi-asynchronous federated learning, Servers, Synchronization, Training, Wireless networks
@article{zheng_over–air_2025,
title = {Over-the-Air Computation Enabled Semi-Asynchronous Wireless Federated Learning},
author = {Zijian Zheng and Yansha Deng and Wenqiang Yi and Hyundong Shin and Arumugam Nallanathan},
url = {https://ieeexplore.ieee.org/document/11048956},
doi = {10.1109/TCOMM.2025.3582727},
issn = {1558-0857},
year = {2025},
date = {2025-01-01},
urldate = {2025-10-08},
journal = {IEEE Transactions on Communications},
pages = {1–1},
abstract = {The emerging field of federated learning (FL) holds significant promise for advancing edge intelligence while preserving data privacy. However, as FL systems scale or become more heterogeneous, challenges such as spectrum scarcity and the straggler problem arise. To address these issues, this paper proposes SA-AirFed, a semi-asynchronous FL architecture compatible with Over-the-Air Computation (AirComp). We develop an efficient scheduling scheme that meets AirComp’s requirements and analyze the factors affecting convergence under the Lipschitz-Smooth condition. Building on insights from the convergence analysis, we design an adaptive algorithm that mitigates staleness from semi-asynchronous aggregation and noise from AirComp by dynamically adjusting aggregation weights, formulated as a convex quadratic programming problem. Experimental results on MNIST and CIFAR-10 demonstrate that SA-AirFed significantly reduces wall-clock training time while achieving greater robustness compared to baseline models.},
keywords = {aggregation optimization, Atmospheric modeling, Computational modeling, Convergence, Federated learning, Noise, Optimization, over-the-air computation, Semi-asynchronous federated learning, Servers, Synchronization, Training, Wireless networks},
pubstate = {published},
tppubtype = {article}
}
The emerging field of federated learning (FL) holds significant promise for advancing edge intelligence while preserving data privacy. However, as FL systems scale or become more heterogeneous, challenges such as spectrum scarcity and the straggler problem arise. To address these issues, this paper proposes SA-AirFed, a semi-asynchronous FL architecture compatible with Over-the-Air Computation (AirComp). We develop an efficient scheduling scheme that meets AirComp’s requirements and analyze the factors affecting convergence under the Lipschitz-Smooth condition. Building on insights from the convergence analysis, we design an adaptive algorithm that mitigates staleness from semi-asynchronous aggregation and noise from AirComp by dynamically adjusting aggregation weights, formulated as a convex quadratic programming problem. Experimental results on MNIST and CIFAR-10 demonstrate that SA-AirFed significantly reduces wall-clock training time while achieving greater robustness compared to baseline models.
2.
Chen, Tianrui; Zhang, Xinruo; You, Minglei; Zheng, Gan; Lambotharan, Sangarapillai
Federated Learning Enabled Link Scheduling in D2D Wireless Networks Journal Article
In: IEEE Wireless Communications Letters, vol. 13, no. 1, pp. 89–92, 2024, ISSN: 2162-2345.
Abstract | Links | BibTeX | Tags: Computational modeling, device-to-device (D2D), Device-to-device communication, Federated learning, link scheduling, Scheduling, Servers, Training, Wireless networks
@article{chen_federated_2024,
title = {Federated Learning Enabled Link Scheduling in D2D Wireless Networks},
author = {Tianrui Chen and Xinruo Zhang and Minglei You and Gan Zheng and Sangarapillai Lambotharan},
url = {https://ieeexplore.ieee.org/document/10268986},
doi = {10.1109/LWC.2023.3321500},
issn = {2162-2345},
year = {2024},
date = {2024-01-01},
urldate = {2025-10-08},
journal = {IEEE Wireless Communications Letters},
volume = {13},
number = {1},
pages = {89–92},
abstract = {Centralized machine learning methods for device-to-device (D2D) link scheduling may lead to a computing burden for a central server, transmission latency for decisions, and privacy issues for D2D communications. To mitigate these challenges, a federated learning (FL) based method is proposed to solve the link scheduling problem, where a global model is distributedly trained at local devices, and a server is used for aggregating model parameters instead of training samples. Specially, a more realistic scenario with limited channel state information (CSI) is considered instead of full CSI. Despite a decentralized implementation, simulation results demonstrate that the proposed FL based approach with limited CSI performs close to the conventional optimization algorithm. In addition, the FL based solution achieves almost the same performance as that of the centralized training.},
keywords = {Computational modeling, device-to-device (D2D), Device-to-device communication, Federated learning, link scheduling, Scheduling, Servers, Training, Wireless networks},
pubstate = {published},
tppubtype = {article}
}
Centralized machine learning methods for device-to-device (D2D) link scheduling may lead to a computing burden for a central server, transmission latency for decisions, and privacy issues for D2D communications. To mitigate these challenges, a federated learning (FL) based method is proposed to solve the link scheduling problem, where a global model is distributedly trained at local devices, and a server is used for aggregating model parameters instead of training samples. Specially, a more realistic scenario with limited channel state information (CSI) is considered instead of full CSI. Despite a decentralized implementation, simulation results demonstrate that the proposed FL based approach with limited CSI performs close to the conventional optimization algorithm. In addition, the FL based solution achieves almost the same performance as that of the centralized training.