Scholarlink

Federated Learning (FL) is developed as a privacy-preserving approach for collaborative model training, initially developed to safeguard sensitive data. While traditionally centralized, FL’s reliance on a central server introduces vulnerabilities and limits user control. To address these issues, decentralized FL has been explored, though it faces challenges in areas including privacy issues, aggregation delays, and difficulties in implementation. To address these concerns, we present a new FL-based approach with a practical decentralized Peer-to-Peer (P2P) system combined with Differential Privacy (DP) to enhance privacy. Our approach uses a decentralized weighted averaging mechanism based on accuracy to improve privacy and model utility. Validation in both virtual and real-world environments demonstrates the effectiveness of our work, and comparative analysis shows an improved performance over existing centralized and decentralized FL systems. Additionally, we demonstrate improvements in KERNEL time, compare model aggregation with and without DP, and examine the impact of our proposed aggregation method on convergence compared to existing methods. A practical application of our protocol in a recommendation system for smart shopping carts highlights its capability to efficiently perform privacy-preserved aggregations in a decentralized network. Our findings suggest significant advancements in FL’s security and performance through this innovative approach.

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

Peer-to-Peer Collaborative Learning Platform for Privacy-Preserved Federated Learning in Industrial Internet

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Kavini Kushani,
  • Isuru Lakshan,
  • Lakshan Pathiraja,
  • Chamara Sandeepa,
  • Tharindu Gamage,
  • Thilina Weerasinghe,
  • Madhusanka Liyanage

摘要

Federated Learning (FL) is developed as a privacy-preserving approach for collaborative model training, initially developed to safeguard sensitive data. While traditionally centralized, FL’s reliance on a central server introduces vulnerabilities and limits user control. To address these issues, decentralized FL has been explored, though it faces challenges in areas including privacy issues, aggregation delays, and difficulties in implementation. To address these concerns, we present a new FL-based approach with a practical decentralized Peer-to-Peer (P2P) system combined with Differential Privacy (DP) to enhance privacy. Our approach uses a decentralized weighted averaging mechanism based on accuracy to improve privacy and model utility. Validation in both virtual and real-world environments demonstrates the effectiveness of our work, and comparative analysis shows an improved performance over existing centralized and decentralized FL systems. Additionally, we demonstrate improvements in KERNEL time, compare model aggregation with and without DP, and examine the impact of our proposed aggregation method on convergence compared to existing methods. A practical application of our protocol in a recommendation system for smart shopping carts highlights its capability to efficiently perform privacy-preserved aggregations in a decentralized network. Our findings suggest significant advancements in FL’s security and performance through this innovative approach.