Scholarlink

<p>Federated Learning (FL) has emerged as a powerful paradigm for modeling and optimizing interactions among distributed systems, enabling collaborative training without direct data sharing. However, practical deployments encounter challenges such as system heterogeneity, unbalanced communication efficiency, and the need for privacy-preserving coordination. Personalized Federated Learning (PFL) mitigates part of these issues by tailoring models to individual systems, yet it often suffers from limited generalization and unstable aggregation in heterogeneous environments. To address these challenges, we propose DP-Pflswar, a novel PFL framework that integrates Differential Privacy (DP) and Stochastic Weight Averaging (SWA) to jointly enhance personalization, generalization, and secure system-to-system communication. The framework introduces a random client-selection strategy for SWA-based aggregation, dynamically combining information from diverse client systems to improve robustness and stability. Moreover, we design a layer-importance–based DP allocation mechanism, which applies noise selectively to shared layers, balancing model performance and privacy protection in the communication process. We provide a non-convex convergence analysis to establish the theoretical soundness of the proposed method and conduct extensive experiments on multiple benchmark datasets. Results demonstrate that DP-Pflswar consistently outperforms existing personalized and differentially private federated learning approaches, offering a hybrid solution that bridges machine learning, communication optimization, and system interaction modeling.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dp-pflswar: a personalized federated learning framework integrating differential privacy and stochastic weight averaging

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • Yifan Mao,
  • Fangfang Shan,
  • Shiqi Mao,
  • Lulu Fan,
  • Zhuo Chen,
  • Yuhang Liu

摘要

Federated Learning (FL) has emerged as a powerful paradigm for modeling and optimizing interactions among distributed systems, enabling collaborative training without direct data sharing. However, practical deployments encounter challenges such as system heterogeneity, unbalanced communication efficiency, and the need for privacy-preserving coordination. Personalized Federated Learning (PFL) mitigates part of these issues by tailoring models to individual systems, yet it often suffers from limited generalization and unstable aggregation in heterogeneous environments. To address these challenges, we propose DP-Pflswar, a novel PFL framework that integrates Differential Privacy (DP) and Stochastic Weight Averaging (SWA) to jointly enhance personalization, generalization, and secure system-to-system communication. The framework introduces a random client-selection strategy for SWA-based aggregation, dynamically combining information from diverse client systems to improve robustness and stability. Moreover, we design a layer-importance–based DP allocation mechanism, which applies noise selectively to shared layers, balancing model performance and privacy protection in the communication process. We provide a non-convex convergence analysis to establish the theoretical soundness of the proposed method and conduct extensive experiments on multiple benchmark datasets. Results demonstrate that DP-Pflswar consistently outperforms existing personalized and differentially private federated learning approaches, offering a hybrid solution that bridges machine learning, communication optimization, and system interaction modeling.