Efficient Personalized Federated Learning via Adaptive Weight Clustering Pruning
Keywords: Peronalized Federated Learning, Communication Efficiency, Data Heterogeneity, Bandwidth Heterogeneity
TL;DR: AdFedWCP is a personalized federated learning method that improves communication efficiency and model accuracy by using adaptive weight clustering pruning to address data and bandwidth heterogeneity in clients.
Abstract: This paper introduces a novel personalized federated learning approach, Adaptive Federated Weight Clustering Pruning (AdFedWCP), specifically designed to optimize communication efficiency in heterogeneous network environments. AdFedWCP innovatively combines adaptive weight clustering pruning techniques, effectively addressing data and bandwidth heterogeneity. By dynamically adjusting clustering centroids based on layer importance and client-specific data characteristics, it significantly reduces communication overhead. Experimental results show that AdFedWCP achieves a reduction in communication volume ranging from 87.54% to 87.82% in communication volume, surpassing the state-of-the-art work on reducing communication overhead in personalized federated learning. AdFedWCP also surpasses existing methods in terms of accuracy across multiple datasets, with improvements ranging from 9.13% to 21.79% over the baselines on EMNIST, CIFAR-10, and CIFAR-100. These results highlight AdFedWCP’s advantages in balancing communication efficiency and model accuracy, making it an ideal choice for resource-constrained federated learning environments.
Primary Area: optimization
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4243
Loading