EUVIMEDEUVIMED entdecken
Crossref

Efficiency-Aware Federated Learning for Image Classification via Model Adaptation

Yiheng Lu, Ziyu Guan, Maoguo Gong, Wei Zhao, Zhuping Hu, Fenlong Jiang

IEEE Transactions on Circuits and Systems for Video Technology · 2026-08 · 36 (8) · 11607-11624

DOI: 10.1109/tcsvt.2026.3680099 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Circuits and Systems for Video Technology, 2026-08. Verfasst von Yiheng Lu und 5 weiteren Autor:innen. Die Arbeit nennt 59 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Key Research and Development Program of Shaanxi, Xidian University Specially Funded Project for Interdisciplinary Exploration.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Massive client training and bandwidth occupation are two main efficiency issues in Federated Learning (FL) for image classification. In this paper, we propose a model adaptation technique to alleviate the efficiency concerns of the FL. The model adaptation technique incorporates two key components: a Spectrum-Tempered Filter Pruning Framework (STFPF) and an Adaptive Accuracy-Based Selection Mechanism (AABSM), which are designed to effectively reduce client-side training costs and minimize bandwidth consumption, respectively. STFPF employs spectral analysis of the weight matrices to dynamically determine the optimal pruning rate. Specifically, the framework analyzes the eigenvalue distribution of each layer’s weight matrix, enabling the identification and removal of filters based on the magnitude of their corresponding eigenvalues. This spectral-based pruning mechanism allows STFPF to effectively reduce the scale of local models, thereby significantly decreasing computational costs throughout the FL process. Moreover, AABSM intelligently selects a subset of clients for model aggregation based on their individual model performance metrics. This performance-driven selection approach significantly reduces the consumption of required communication time between the central server and local clients, thereby effectively conserving bandwidth resources throughout the FL process. We extensively evaluate our model adaptation technique on three prominent architectures with AlexNet, VGG-16, and ResNet-50, using the FashionMNIST, CIFAR-10, and CIFAR-100 datasets, respectively. For AlexNet, out approach can reduce the computation cost obviously while with tiny accuracy sacrifice. For VGG-16 and ResNet-50, our approach not only achieves superior accuracy in the central model compared to baseline methods but also yields significant reductions in computational complexity (FLOPs), model parameters, and bandwidth consumption throughout the complete federated learning cycle.

Förderung

  • National Natural Science Foundation of China
  • Key Research and Development Program of Shaanxi
  • Xidian University Specially Funded Project for Interdisciplinary Exploration

Zitieren

APA

Lu, Y., Guan, Z., Gong, M., Zhao, W., Hu, Z., & Jiang, F. (2026). Efficiency-Aware Federated Learning for Image Classification via Model Adaptation. IEEE Transactions on Circuits and Systems for Video Technology, 36(8), 11607-11624. https://doi.org/10.1109/tcsvt.2026.3680099

Vancouver

Lu Y, Guan Z, Gong M, Zhao W, Hu Z, Jiang F. Efficiency-Aware Federated Learning for Image Classification via Model Adaptation. IEEE Transactions on Circuits and Systems for Video Technology. 2026;36(8):11607-11624. doi:10.1109/tcsvt.2026.3680099

BibTeX
@article{Lu2026,
  author = {Lu, Yiheng and Guan, Ziyu and Gong, Maoguo and Zhao, Wei and Hu, Zhuping and Jiang, Fenlong},
  title = {Efficiency-Aware Federated Learning for Image Classification via Model Adaptation},
  journal = {IEEE Transactions on Circuits and Systems for Video Technology},
  year = {2026},
  volume = {36},
  number = {8},
  pages = {11607--11624},
  doi = {10.1109/tcsvt.2026.3680099}
}
Kostenlos starten →