Scholarlink

<p>As a fundamental method in the field of machine learning, federated learning is based on the principle of protecting the confidentiality of client data by requiring only the transfer of local models from clients to a central server, thereby avoiding the leakage of the underlying data. However, recent research suggests that even the exchange of model parameters between clients can indirectly reveal sensitive information about the local training data sets. Consequently, homomorphic encryption, which provides verifiable security without increasing communication overhead, has become one of the most commonly used methods. However, existing FL solutions using homomorphic encryption typically assume that all participating clients uniformly use public-private key pairs, which may inadvertently lead to the leakage of model gradients and thus compromise the confidentiality of client data. To address these issues, this paper proposes EMKPPA, an efficient multi-key privacy-preserving aggregation scheme specifically designed for FL, which aims to ensure secure and effective aggregation within a multi-key framework. The system adopts a dual-server architecture and exploits proxy re-encryption technology, combined with the additive properties of homomorphic encryption and the secure key exchange protocol of Diffie-Hellman, to achieve robust ciphertext transformation and aggregation processes, thereby ensuring the confidentiality of client data. Comprehensive security analysis and performance evaluation on various datasets provide strong evidence of the effectiveness and efficiency of EMKPPA in multi-key aggregation 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

EMKPPA: Efficient multi-key privacy-preserving aggregation in federated learning

  • 期刊论文

  • 发表日期：

    2025年7月2日

查看全文
  • Zhen Liu,
  • Changsong Yang,
  • Yong Ding,
  • Yujue Wang,
  • Huiyong Wang

摘要

As a fundamental method in the field of machine learning, federated learning is based on the principle of protecting the confidentiality of client data by requiring only the transfer of local models from clients to a central server, thereby avoiding the leakage of the underlying data. However, recent research suggests that even the exchange of model parameters between clients can indirectly reveal sensitive information about the local training data sets. Consequently, homomorphic encryption, which provides verifiable security without increasing communication overhead, has become one of the most commonly used methods. However, existing FL solutions using homomorphic encryption typically assume that all participating clients uniformly use public-private key pairs, which may inadvertently lead to the leakage of model gradients and thus compromise the confidentiality of client data. To address these issues, this paper proposes EMKPPA, an efficient multi-key privacy-preserving aggregation scheme specifically designed for FL, which aims to ensure secure and effective aggregation within a multi-key framework. The system adopts a dual-server architecture and exploits proxy re-encryption technology, combined with the additive properties of homomorphic encryption and the secure key exchange protocol of Diffie-Hellman, to achieve robust ciphertext transformation and aggregation processes, thereby ensuring the confidentiality of client data. Comprehensive security analysis and performance evaluation on various datasets provide strong evidence of the effectiveness and efficiency of EMKPPA in multi-key aggregation scenarios.