Scholarlink

<p>As an emerging distributed machine learning paradigm, federated learning enables collaborative model training across multiple participants while safeguarding local data privacy. However, the development of federated learning on a central server introduces critical vulnerabilities, the single-point-of-failure risk. To address the current limitations of federated learning, this paper proposes an innovative practical Byzantine fault tolerance (RP-PBFT) consensus mechanism with reputation evaluation and proxy signature. Moreover, this paper proposed a comprehensive reputation evaluation model that dynamically assesses node trustworthiness through two critical dimensions, which one is the deviation between locally loaded gradients and global gradients, the other is the behavior of the nodes during the consensus process. Furthermore, the proposed mechanism integrates a (<i>t</i>,&#xa0;<i>n</i>) threshold proxy signature to delegate consensus participation rights for suspicious nodes, thereby improving the system’s effective fault tolerance under proxy-assisted consensus participation. Experimental results demonstrate that RP-PBFT reduces communication overhead by over 40% compared with traditional PBFT, while preserving consensus consistency with high malicious-node ratios through proxy-assisted reduction of effective Byzantine participation. By leveraging reputation-weighted primary node election and gradient anomaly detection, RP-PBFT effectively mitigates gradient poisoning, sybil attacks and collusion attacks.</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

RP-PBFT: a fault-tolerance-enhanced PBFT consensus with reputation evaluation for federated learning

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Yong Zhang,
  • Tong Chen,
  • Dengzhi Liu,
  • Zhen Zhang

摘要

As an emerging distributed machine learning paradigm, federated learning enables collaborative model training across multiple participants while safeguarding local data privacy. However, the development of federated learning on a central server introduces critical vulnerabilities, the single-point-of-failure risk. To address the current limitations of federated learning, this paper proposes an innovative practical Byzantine fault tolerance (RP-PBFT) consensus mechanism with reputation evaluation and proxy signature. Moreover, this paper proposed a comprehensive reputation evaluation model that dynamically assesses node trustworthiness through two critical dimensions, which one is the deviation between locally loaded gradients and global gradients, the other is the behavior of the nodes during the consensus process. Furthermore, the proposed mechanism integrates a (t, n) threshold proxy signature to delegate consensus participation rights for suspicious nodes, thereby improving the system’s effective fault tolerance under proxy-assisted consensus participation. Experimental results demonstrate that RP-PBFT reduces communication overhead by over 40% compared with traditional PBFT, while preserving consensus consistency with high malicious-node ratios through proxy-assisted reduction of effective Byzantine participation. By leveraging reputation-weighted primary node election and gradient anomaly detection, RP-PBFT effectively mitigates gradient poisoning, sybil attacks and collusion attacks.