Scholarlink

<p>Federated learning attained significant worldwide recognition due to the capability to provide distributed training of deep learning models on various devices while ensuring user data privacy. Federated Learning revolutionizes machine learning greatly. However, technology, non-independent and identically distributed (non-IID) data, communication overhead, and privacy problems hinder its deployment. Recent studies have also shown that training models on non-IID data reduces the performance, convergence rate, and global model quality in federated learning. Current federated learning and clustered federated learning algorithms face two key issues: ineffective training of the clients and static utilization of hyperparameters. This paper presents a new algorithm named Real-Coded Genetic Clustered Federated Learning (<i>ReGen-CFL</i>) to overcome the above key issues. This algorithm uses a density-based clustering technique named Ordering Points To Identify the Clustering Structure (OPTICS) for clustering the client devices on the basis of training hyperparameters. It uses real-coded genetic optimization after forming the clusters to genetically modify and evolve the hyperparameters for every cluster to fit each cluster’s specific requirements better. It is mainly proposed for overcoming the limitation of static utilization of hyperparameters to fine-tune the model, improving the model performance and enhancing the communication efficiency in clustered federated learning. The effectiveness of the proposed algorithm is showcased by performing comprehensive experiments on benchmark datasets. Experimental findings confirm the effectiveness of the <i>ReGen-CFL</i> algorithm, attaining an accuracy of 79.12% on the CIFAR-10 dataset, 99.38% on MNIST, and 92.54% on Fashion-MNIST. The findings reveal that <i>ReGen-CFL</i> performs better than conventional federated learning and clustered federated learning approaches, even in cases where the underlying clustering structure is not readily observable or where non-IID data exists.</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

Ameliorating clustered federated learning using real-coded genetic algorithm

  • 期刊论文

  • 发表日期：

    2025年7月30日

查看全文
  • Rajesh Kumar Chaudhary,
  • Ravinder Kumar,
  • Nitin Saxena

摘要

Federated learning attained significant worldwide recognition due to the capability to provide distributed training of deep learning models on various devices while ensuring user data privacy. Federated Learning revolutionizes machine learning greatly. However, technology, non-independent and identically distributed (non-IID) data, communication overhead, and privacy problems hinder its deployment. Recent studies have also shown that training models on non-IID data reduces the performance, convergence rate, and global model quality in federated learning. Current federated learning and clustered federated learning algorithms face two key issues: ineffective training of the clients and static utilization of hyperparameters. This paper presents a new algorithm named Real-Coded Genetic Clustered Federated Learning (ReGen-CFL) to overcome the above key issues. This algorithm uses a density-based clustering technique named Ordering Points To Identify the Clustering Structure (OPTICS) for clustering the client devices on the basis of training hyperparameters. It uses real-coded genetic optimization after forming the clusters to genetically modify and evolve the hyperparameters for every cluster to fit each cluster’s specific requirements better. It is mainly proposed for overcoming the limitation of static utilization of hyperparameters to fine-tune the model, improving the model performance and enhancing the communication efficiency in clustered federated learning. The effectiveness of the proposed algorithm is showcased by performing comprehensive experiments on benchmark datasets. Experimental findings confirm the effectiveness of the ReGen-CFL algorithm, attaining an accuracy of 79.12% on the CIFAR-10 dataset, 99.38% on MNIST, and 92.54% on Fashion-MNIST. The findings reveal that ReGen-CFL performs better than conventional federated learning and clustered federated learning approaches, even in cases where the underlying clustering structure is not readily observable or where non-IID data exists.