Scholarlink

<p>Federated learning (FL) enables decentralized training of machine learning models across distributed Internet of Things (IoT) clients while preserving data locality. However, achieving strong differential privacy (DP) guarantees under heterogeneous, non-IID data remains a critical challenge. This paper presents Div-DPFL, a divergence-aware, privacy-preserving framework that adaptively calibrates client-level noise based on statistical divergence, sensitivity, and participation frequency. Heterogeneity is formally modeled through weighted Rényi divergence to capture discrepancies between local and global data distributions, with privacy loss analyzed under zero-concentrated differential privacy (zCDP). The proposed mechanism scales Gaussian noise according to divergence-based sensitivity, enabling tighter zCDP composition and personalized privacy budgeting. To meet the high computational demands of large-scale federated training, Div-DPFL is designed for parallel and distributed execution on high-performance computing (HPC) infrastructures, leveraging GPU acceleration for real-time gradient aggregation and adaptive noise scaling. Extensive experiments on CIFAR-10, FEMNIST, and synthetic IoT datasets demonstrate that Div-DPFL improves accuracy by up to 6.8% and reduces average privacy loss by 11.5% compared with state-of-the-art DP-FL methods. It also enhances fairness by minimizing variance in client-level accuracy and privacy cost. Notably, Div-DPFL sustains stable performance under extreme heterogeneity (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\alpha = 0.1\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>α</mi> <mo>=</mo> <mn>0.1</mn> </mrow> </math></EquationSource> </InlineEquation>) and adapts effectively to varying privacy budgets and participation rates. Overall, Div-DPFL establishes a scalable and computation-aware foundation for privacy-preserving federated learning in heterogeneous IoT environments, aligning with the real-time and parallel processing paradigms of modern supercomputing.</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

Mathematical bounds on differential privacy in heterogeneous federated IoT environments

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • Waseem Abbass,
  • Nasim Abbas,
  • Uzma Majeed

摘要

Federated learning (FL) enables decentralized training of machine learning models across distributed Internet of Things (IoT) clients while preserving data locality. However, achieving strong differential privacy (DP) guarantees under heterogeneous, non-IID data remains a critical challenge. This paper presents Div-DPFL, a divergence-aware, privacy-preserving framework that adaptively calibrates client-level noise based on statistical divergence, sensitivity, and participation frequency. Heterogeneity is formally modeled through weighted Rényi divergence to capture discrepancies between local and global data distributions, with privacy loss analyzed under zero-concentrated differential privacy (zCDP). The proposed mechanism scales Gaussian noise according to divergence-based sensitivity, enabling tighter zCDP composition and personalized privacy budgeting. To meet the high computational demands of large-scale federated training, Div-DPFL is designed for parallel and distributed execution on high-performance computing (HPC) infrastructures, leveraging GPU acceleration for real-time gradient aggregation and adaptive noise scaling. Extensive experiments on CIFAR-10, FEMNIST, and synthetic IoT datasets demonstrate that Div-DPFL improves accuracy by up to 6.8% and reduces average privacy loss by 11.5% compared with state-of-the-art DP-FL methods. It also enhances fairness by minimizing variance in client-level accuracy and privacy cost. Notably, Div-DPFL sustains stable performance under extreme heterogeneity ( \(\alpha = 0.1\) α = 0.1 ) and adapts effectively to varying privacy budgets and participation rates. Overall, Div-DPFL establishes a scalable and computation-aware foundation for privacy-preserving federated learning in heterogeneous IoT environments, aligning with the real-time and parallel processing paradigms of modern supercomputing.