Adaptive Sparse Ternary Compression with Dynamic Gradient Thresholding for Communication-Efficient Federated Learning
Engineering, Technology & Applied Science Research · 06.06.2026 · 16 (3) · 35281-35286
DOI: 10.48084/etasr.18548 ↗Kurz gefasst
Fachartikel in der Zeitschrift Engineering, Technology & Applied Science Research, online erschienen am 06.06.2026. Verfasst von Nithyaniranjana Murthy Chittaiah und 1 weiteren Autor:innen. Die Arbeit nennt 25 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Federated Learning (FL) supports collaborative model training while not exchanging raw data, but suffers from scalability issues due to a significant communication overhead when updating models too frequently. Sparse Ternary Compression (STC) addresses this by both gradient sparsification and ternarization; however, fixed sparsification thresholds do not adapt to gradually changing gradient distributions during training. This study introduces a dynamic gradient sparsification method, called Adaptive-STC, which learns optimized sparsification thresholds for each communication round by using statistics from local gradients. The proposed strategy provides an adaptive mechanism to control sparsity during training and results in better communication efficiency while maintaining the model's accuracy. Experimental results on both CIFAR-10 and MedMNIST using a lightweight version of VGG11 show that Adaptive-STC reduces communication cost by up to 18% compared to the dense FedAvg and fixed-threshold STC, while incurring less than 0.3% performance degradation. These findings demonstrate the value of adaptive thresholding in data-efficient FL.
Zitieren
APA
Chittaiah, N. M., & Manjula, S. H. (2026). Adaptive Sparse Ternary Compression with Dynamic Gradient Thresholding for Communication-Efficient Federated Learning. Engineering, Technology & Applied Science Research, 16(3), 35281-35286. https://doi.org/10.48084/etasr.18548
Vancouver
Chittaiah NM, Manjula SH. Adaptive Sparse Ternary Compression with Dynamic Gradient Thresholding for Communication-Efficient Federated Learning. Engineering, Technology & Applied Science Research. 2026;16(3):35281-35286. doi:10.48084/etasr.18548
BibTeX
@article{Chittaiah2026,
author = {Chittaiah, Nithyaniranjana Murthy and Manjula, S. H.},
title = {Adaptive Sparse Ternary Compression with Dynamic Gradient Thresholding for Communication-Efficient Federated Learning},
journal = {Engineering, Technology & Applied Science Research},
year = {2026},
volume = {16},
number = {3},
pages = {35281--35286},
doi = {10.48084/etasr.18548}
}
EUVIMED