Scholarlink

Federated Learning (FL) is a decentralized machine learning approach that enables model training over distributed nodes without sharing the private data. However, FL faces various challenges, such as limited scalability, handling data heterogeneity, and the impact of malicious clients on model performance. In this paper, we propose a hierarchical FL framework using gossip protocol and Differential Privacy Mondrian (DPM) clustering. Our framework follows a three-level architecture that consists of a central server, cluster heads, and data owners to ensure scalability. The proposed framework addresses data heterogeneity by grouping clients based on statistical properties of data. It also assigns weights to clients based on their contribution to the global model, thereby mitigating the impact of malicious clients. To evaluate the performance of our proposed framework, we experimented on three well-known FL datasets, MNIST, FMNIST, and CIFAR-10. The results show that our framework achieves 2.1% higher accuracy, reduces training loss by 15.46%, and has less execution time compared to the standard FedAvg algorithm. The results demonstrate that the framework performs better in terms of training efficiency and model accuracy while preserving privacy.

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

A Secure Federated Learning Using Differential Privacy Mondrian Clustering

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Rojalini Tripathy,
  • Paladri Pranitha,
  • B. U. Tejonath,
  • Padmalochan Bera

摘要

Federated Learning (FL) is a decentralized machine learning approach that enables model training over distributed nodes without sharing the private data. However, FL faces various challenges, such as limited scalability, handling data heterogeneity, and the impact of malicious clients on model performance. In this paper, we propose a hierarchical FL framework using gossip protocol and Differential Privacy Mondrian (DPM) clustering. Our framework follows a three-level architecture that consists of a central server, cluster heads, and data owners to ensure scalability. The proposed framework addresses data heterogeneity by grouping clients based on statistical properties of data. It also assigns weights to clients based on their contribution to the global model, thereby mitigating the impact of malicious clients. To evaluate the performance of our proposed framework, we experimented on three well-known FL datasets, MNIST, FMNIST, and CIFAR-10. The results show that our framework achieves 2.1% higher accuracy, reduces training loss by 15.46%, and has less execution time compared to the standard FedAvg algorithm. The results demonstrate that the framework performs better in terms of training efficiency and model accuracy while preserving privacy.