Scholarlink

This study addresses the issue of finding the desired balance between privacy and accuracy of privacy-preserving predictive model in decentralized healthcare environments by integrating federated learning with differential privacy mechanism while automating the process of implementation of desired privacy budget selection using novel epsilon-aware strategy. The analysis compares different machine learning models like EfficientNet, ResNet, DenseNet, and MobileNet, with EfficientNet achieving the highest accuracy of 84.58%. In differential privacy, Epsilon determines the privacy loss limit where a smaller epsilon value guarantees stricter privacy by limiting the individual data point’s influence on model outputs. However, overly restrictive epsilon degrades the model’s usefulness due to excessive noise injection. The study focuses on dynamically selecting the best epsilon based on given privacy-performance requirements in differential privacy with theoretical and empirical validation to find the desired balance between privacy and accuracy in the federated learning environment.

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

Automating the Optimal Privacy Budget Selection for Differential Privacy in Federated Learning Environments

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Faisal Mahmud,
  • Abdul Mueed,
  • Mohammad Taseen,
  • Md Jahir Uddin Shuvo,
  • Nusrat Jahan Pritu,
  • Rashedur M. Rahman

摘要

This study addresses the issue of finding the desired balance between privacy and accuracy of privacy-preserving predictive model in decentralized healthcare environments by integrating federated learning with differential privacy mechanism while automating the process of implementation of desired privacy budget selection using novel epsilon-aware strategy. The analysis compares different machine learning models like EfficientNet, ResNet, DenseNet, and MobileNet, with EfficientNet achieving the highest accuracy of 84.58%. In differential privacy, Epsilon determines the privacy loss limit where a smaller epsilon value guarantees stricter privacy by limiting the individual data point’s influence on model outputs. However, overly restrictive epsilon degrades the model’s usefulness due to excessive noise injection. The study focuses on dynamically selecting the best epsilon based on given privacy-performance requirements in differential privacy with theoretical and empirical validation to find the desired balance between privacy and accuracy in the federated learning environment.