Scholarlink

Federated Learning (FL) is a distributed machine learning approach that protects data privacy by enabling multiple devices or nodes to collaboratively train models without sharing raw data. However, existing research highlights significant challenges in terms of privacy protection and communication efficiency. Gradient leakage remains a key issue in FL security, while traditional FL frameworks incur a large amount of communication overhead in large-scale scenarios, limiting their applicability. To address these issues, this study proposes GroupRingFL, a secure and efficient ring-aggregation federated learning method. This framework dynamically adjusts the number of groups to optimize communication overhead in large-scale federated learning scenarios. Additionally, an efficient double-masking mechanism is applied within this framework, effectively defending against collusion attacks by honest-but-curious adversaries and addressing user dropout problems. Extensive experiments on the MNIST and CIFAR-100 datasets using MLP, CNN, and L-BFGS models demonstrate that GroupRingFL achieves high accuracy comparable to federated averaging, while offering superior security compared to differential privacy-based FL and better communication efficiency and storage requirements than traditional secure aggregation methods.

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

Robust and Efficient Group-Based Ring Federated Learning Framework with Double-Masking Mechanism

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Changji Wang,
  • Boxuan Lin,
  • Qingqing Gan,
  • Ning Liu,
  • Zhen Liu

摘要

Federated Learning (FL) is a distributed machine learning approach that protects data privacy by enabling multiple devices or nodes to collaboratively train models without sharing raw data. However, existing research highlights significant challenges in terms of privacy protection and communication efficiency. Gradient leakage remains a key issue in FL security, while traditional FL frameworks incur a large amount of communication overhead in large-scale scenarios, limiting their applicability. To address these issues, this study proposes GroupRingFL, a secure and efficient ring-aggregation federated learning method. This framework dynamically adjusts the number of groups to optimize communication overhead in large-scale federated learning scenarios. Additionally, an efficient double-masking mechanism is applied within this framework, effectively defending against collusion attacks by honest-but-curious adversaries and addressing user dropout problems. Extensive experiments on the MNIST and CIFAR-100 datasets using MLP, CNN, and L-BFGS models demonstrate that GroupRingFL achieves high accuracy comparable to federated averaging, while offering superior security compared to differential privacy-based FL and better communication efficiency and storage requirements than traditional secure aggregation methods.