Scholarlink

<p>The progression of machine learning into distributed paradigms, such as decentralized federated learning (DFL), presents a novel approach to handling data across multiple clients without centralization, thereby enhancing data privacy and system efficiency. This study assessed the impact of network topologies on DFL performance, particularly in terms of accuracy and communication overhead. Utilizing the CIFAR-10 and SST-5 datasets, our experiments evaluated the effects of star, binary tree, random, k-connected, and fully connected topologies on key metrics such as model accuracy and bandwidth utilization. Our analysis revealed trade-offs in network topology selection: higher connectivity topologies enhance learning outcomes in independent and identically distributed (IID) data scenarios but come with increased bandwidth requirements, whereas lower-degree topologies can balance performance with communication costs under non-IID conditions. These insights emphasize the importance of selecting network topologies to develop secure, scalable, and efficient systems, and offer valuable guidance for optimizing decentralized learning frameworks in evolving data landscapes.</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

Impact of network topologies on decentralized federated learning

  • 期刊论文

  • 发表日期：

    2025年10月9日

查看全文
  • Radwan Selo,
  • Majid Kundroo,
  • Taehong Kim

摘要

The progression of machine learning into distributed paradigms, such as decentralized federated learning (DFL), presents a novel approach to handling data across multiple clients without centralization, thereby enhancing data privacy and system efficiency. This study assessed the impact of network topologies on DFL performance, particularly in terms of accuracy and communication overhead. Utilizing the CIFAR-10 and SST-5 datasets, our experiments evaluated the effects of star, binary tree, random, k-connected, and fully connected topologies on key metrics such as model accuracy and bandwidth utilization. Our analysis revealed trade-offs in network topology selection: higher connectivity topologies enhance learning outcomes in independent and identically distributed (IID) data scenarios but come with increased bandwidth requirements, whereas lower-degree topologies can balance performance with communication costs under non-IID conditions. These insights emphasize the importance of selecting network topologies to develop secure, scalable, and efficient systems, and offer valuable guidance for optimizing decentralized learning frameworks in evolving data landscapes.