Differentially Private Personalized Federated Learning Based on Parameter Partitioning
IEEE Transactions on Cognitive Communications and Networking · 2026 · 12 · 7474-7489
DOI: 10.1109/tccn.2026.3685398 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Cognitive Communications and Networking, 2026. Verfasst von Mingyu Pang und 3 weiteren Autor:innen. Die Arbeit nennt 40 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Major Research Plan of Inner Mongolia Natural Science Foundation, Program for Young Talents of Science and Technology in Universities of Inner Mongolia Autonomous Region.
Abstract (Originalsprache)
Differentially Private Personalized Federated Learning (DP-PFL) aims to address client data heterogeneity while providing formal privacy guarantees; however, most existing methods rely on static parameter partitioning and uniform privacy protection, ignoring parameter-level differences in personalization relevance and sensitivity, which often leads to excessive noise injection, unstable optimization, and degraded performance under heterogeneous (non-IID) data distributions. To overcome these limitations, we propose a posterior-inspired personalized federated learning framework that jointly models parameter uncertainty, personalization structure, and differential privacy. By leveraging temporal parameter dynamics, we construct an uncertainty-aware posterior surrogate that enables dynamic partitioning of model parameters into shared and personalized parameter subsets, without performing explicit Bayesian posterior inference. Based on this partitioning, we further introduce an adaptive clipping mechanism and a differentiated differential privacy strategy that assigns distinct noise levels to shared and personalized parameters under a single client-level privacy budget according to their sensitivity profiles, achieving a fine-grained privacy–utility trade-off. We provide a formal privacy analysis and convergence guarantee, showing that parameter partitioning with differentiated noise allocation under a single client-level privacy budget yields a tighter noise-induced error bound than uniform DP mechanisms. Extensive experiments on FEMNIST, SVHN, MNIST, and CIFAR-10 demonstrate that the proposed approach consistently achieves higher accuracy, faster convergence, and stronger robustness than state-of-the-art DP-PFL baselines, particularly in highly non-IID settings.
Förderung
- National Natural Science Foundation of China
- Major Research Plan of Inner Mongolia Natural Science Foundation
- Program for Young Talents of Science and Technology in Universities of Inner Mongolia Autonomous Region
Zitieren
APA
Pang, M., Gao, Y., Liu, Y., & Cui, H. (2026). Differentially Private Personalized Federated Learning Based on Parameter Partitioning. IEEE Transactions on Cognitive Communications and Networking, 12, 7474-7489. https://doi.org/10.1109/tccn.2026.3685398
Vancouver
Pang M, Gao Y, Liu Y, Cui H. Differentially Private Personalized Federated Learning Based on Parameter Partitioning. IEEE Transactions on Cognitive Communications and Networking. 2026;12:7474-7489. doi:10.1109/tccn.2026.3685398
BibTeX
@article{Pang2026,
author = {Pang, Mingyu and Gao, Yongqiang and Liu, Yongmei and Cui, Hualong},
title = {Differentially Private Personalized Federated Learning Based on Parameter Partitioning},
journal = {IEEE Transactions on Cognitive Communications and Networking},
year = {2026},
volume = {12},
pages = {7474--7489},
doi = {10.1109/tccn.2026.3685398}
}
EUVIMED