Scholarlink

<p>In data-driven environment, institutions need to analyze and exchange sensitive information, including medical records, in a secure manner to identify and analyze novel clinical information. Existing machine learning techniques, specifically in the healthcare domain, are centered around a single, disparate data source which fails to address secure, decentralized collaborative computation. This has created a need for secure computing environments that proficiently manage sensitive information while providing real-world collaboration opportunities for different stakeholders.This research introduced a decentralized Federated Continual Learning (FCL) framework that allows institutions to process diabetic data collaboratively and securely without ever sharing actual patient records. Apart from enabling collaboration across locations, this approach helps uncover better predictive patterns and improves model efficiency on the client side. It combines continual learning for good data adaptation, FedAvg (Federated Averaging) for server-side aggregation, TabNet for high local model efficiency and Differential Privacy (DP) for High Security. Experimental results demonstrate that the proposed framework achieves high prediction performance, low forgetting rates during continual learning, and moderate communication overhead. It thus clearly shows the ability of the framework to provide scalability, generalization, and adequate privacy protection while ensuring good prediction performance in real-life healthcare scenarios.</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 Novel Framework for Differential Privacy based Federated Continual Learning for Dynamic Medical Data Analysis

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • S. Pandikumar,
  • V. Pandimurugan,
  • A. Rama Prasath,
  • Sajeer Karattil,
  • Subash Chandra Bose Jaganathan,
  • P. Vigneshwaran

摘要

In data-driven environment, institutions need to analyze and exchange sensitive information, including medical records, in a secure manner to identify and analyze novel clinical information. Existing machine learning techniques, specifically in the healthcare domain, are centered around a single, disparate data source which fails to address secure, decentralized collaborative computation. This has created a need for secure computing environments that proficiently manage sensitive information while providing real-world collaboration opportunities for different stakeholders.This research introduced a decentralized Federated Continual Learning (FCL) framework that allows institutions to process diabetic data collaboratively and securely without ever sharing actual patient records. Apart from enabling collaboration across locations, this approach helps uncover better predictive patterns and improves model efficiency on the client side. It combines continual learning for good data adaptation, FedAvg (Federated Averaging) for server-side aggregation, TabNet for high local model efficiency and Differential Privacy (DP) for High Security. Experimental results demonstrate that the proposed framework achieves high prediction performance, low forgetting rates during continual learning, and moderate communication overhead. It thus clearly shows the ability of the framework to provide scalability, generalization, and adequate privacy protection while ensuring good prediction performance in real-life healthcare scenarios.