Scholarlink

<p>Lung cancer detection in medical imaging presents significant challenges, particularly in preserving patient data privacy while enabling collaborative learning across institutions. Federated Learning (FL) offers a solution by allowing multiple clients to collaboratively train a shared model without exposing private data and leveraging diverse datasets. However, conventional FL algorithms such as FedAvg and FedProx often suffer from issues like client drift and limited adaptability in non-IID settings, leading to suboptimal convergence and degraded performance. To address these limitations, this paper proposes EFedProx, an enhanced FL framework that integrates Multilevel Differential Privacy (MDP) and blockchain-based verification to ensure both privacy protection and model integrity. EFedProx introduces an adaptive proximal term that dynamically adjusts based on client divergence, mitigating client drift and improving model aggregation in non-IID settings. Additionally, MDP dynamically adjusts the noise level to balance privacy and utility, guided by client-specific factors such as data sensitivity (captured through label imbalance, dataset size, and gradient magnitude) and training progress, implemented via a round-wise global decay schedule. To enhance security and trust, EFedProx leverages blockchain-based model verification to detect and reject invalid or tampered updates and ensure tamper-proof aggregation. Decentralized storage using IPFS and MongoDB reduces reliance on centralized servers while maintaining scalable and efficient model management. Experimental evaluations on the LUNA16 and IQ-OTH/NCCD lung cancer datasets demonstrate that EFedProx achieves 4–9% and 3–9% higher accuracy respectively, along with faster convergence under both IID and Non-IID settings compared to baseline FL methods.</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

A privacy-preserving federated learning approach with multilevel differential privacy for lung cancer detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Khadija Begum,
  • Md Mamunur Rashid,
  • Md Ariful Islam Mozumder,
  • Hee Cheol Kim,
  • Moon-Il Joo

摘要

Lung cancer detection in medical imaging presents significant challenges, particularly in preserving patient data privacy while enabling collaborative learning across institutions. Federated Learning (FL) offers a solution by allowing multiple clients to collaboratively train a shared model without exposing private data and leveraging diverse datasets. However, conventional FL algorithms such as FedAvg and FedProx often suffer from issues like client drift and limited adaptability in non-IID settings, leading to suboptimal convergence and degraded performance. To address these limitations, this paper proposes EFedProx, an enhanced FL framework that integrates Multilevel Differential Privacy (MDP) and blockchain-based verification to ensure both privacy protection and model integrity. EFedProx introduces an adaptive proximal term that dynamically adjusts based on client divergence, mitigating client drift and improving model aggregation in non-IID settings. Additionally, MDP dynamically adjusts the noise level to balance privacy and utility, guided by client-specific factors such as data sensitivity (captured through label imbalance, dataset size, and gradient magnitude) and training progress, implemented via a round-wise global decay schedule. To enhance security and trust, EFedProx leverages blockchain-based model verification to detect and reject invalid or tampered updates and ensure tamper-proof aggregation. Decentralized storage using IPFS and MongoDB reduces reliance on centralized servers while maintaining scalable and efficient model management. Experimental evaluations on the LUNA16 and IQ-OTH/NCCD lung cancer datasets demonstrate that EFedProx achieves 4–9% and 3–9% higher accuracy respectively, along with faster convergence under both IID and Non-IID settings compared to baseline FL methods.