Scholarlink

<p>In response to potential user privacy issues that may arise during machine learning processes, federated learning, as the first collaborative and online learning solution without requiring users to upload real data and only uploading model updates, has garnered widespread attention and research. However, model updates that users train locally and upload can still potentially contain sensitive information, raising new privacy protection concerns. Additionally, the need for users to perform extensive local training introduces computational and communication overhead, posing a challenge that necessitates the development of a lightweight federated learning architecture. Considering further privacy requirements, this paper employs a federated learning framework with differential privacy mechanisms. Moreover, it introduces for the first time a Dropout mechanism based on the Fisher information matrix Fisher Dropout to enhance the balance between privacy and efficiency in federated learning. Extensive experiments on real-world datasets validate the effectiveness of the proposed solution. The experimental results demonstrate that, compared to other federated learning methods, this approach exhibits significant advantages in balancing privacy protection and usability.</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

Gradient descent-based lightweight federated learning model for differential privacy-preserving

  • 期刊论文

  • 发表日期：

    2025年8月14日

查看全文
  • Haewon Byeon,
  • Azzah AlGhamdi,
  • Ismail Keshta,
  • Mukesh Soni,
  • Achyut Shankar,
  • Rupesh Gupta

摘要

In response to potential user privacy issues that may arise during machine learning processes, federated learning, as the first collaborative and online learning solution without requiring users to upload real data and only uploading model updates, has garnered widespread attention and research. However, model updates that users train locally and upload can still potentially contain sensitive information, raising new privacy protection concerns. Additionally, the need for users to perform extensive local training introduces computational and communication overhead, posing a challenge that necessitates the development of a lightweight federated learning architecture. Considering further privacy requirements, this paper employs a federated learning framework with differential privacy mechanisms. Moreover, it introduces for the first time a Dropout mechanism based on the Fisher information matrix Fisher Dropout to enhance the balance between privacy and efficiency in federated learning. Extensive experiments on real-world datasets validate the effectiveness of the proposed solution. The experimental results demonstrate that, compared to other federated learning methods, this approach exhibits significant advantages in balancing privacy protection and usability.