Scholarlink

<p>In the federated learning, instead of training local models on individual servers, a global model is generated on the central server by aggregating model parameters transmitted from edge servers. The global model is then sent back to the edge servers. This machine learning approach prioritizes privacy, as no user data is transmitted. The federated learning is steered by the learning protocol. In the synchronous learning protocol, the learning round time is the same for all edge servers, such that all servers can progress synchronously. However, this protocol suffers from the “straggler effect”, where stronger servers will remain idle until weaker servers’ complete one training round. Thus, the asynchronous learning protocol has been introduced as a solution, which enables each server to train at its own pace by relaxing the learning round time. However, this protocol has three drawbacks, as: high communication costs, staleness of some local models, and negative impacts of outdated local models on the global model. To address the drawbacks of both synchronous and asynchronous protocols, the semi-synchronous learning protocol has been proposed. However, existing semi-synchronous protocols are still faced with challenges in fully addressing the above drawbacks. In decentralized topologies, peer-to-peer model exchange allows participating edge servers to continuously update and refine their local models, improving local consistency and reducing dependence on a central coordinator. At the same time, centralized aggregation mechanisms can facilitate global model synchronization and help maintain overall convergence across the network. In this study, a new semi-synchronous federated learning protocol, named SSFL, is presented, which trains a global model by taking all servers into account and incorporating a decentralized topology to balance the accuracy of local models. Additionally, it uses a new client selection mechanism to ensure that the information from all servers is utilized for training the global model. Experiments using the MNIST, Boston Housing, KDD Cup '99, and CIFAR10 datasets demonstrate that the proposed SSFL protocol outperforms existing federated learning protocols in terms of efficient use of all servers’ capacity and information, higher convergence rate of the global model, and lower discrepancy between local models’ accuracy.</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

Semi-synchronous federated learning in edge computing using a semi-decentralized topology

  • 期刊论文

  • 发表日期：

    2026年6月16日

查看全文
  • Mahyar Hassanpour,
  • Reza Ramezani

摘要

In the federated learning, instead of training local models on individual servers, a global model is generated on the central server by aggregating model parameters transmitted from edge servers. The global model is then sent back to the edge servers. This machine learning approach prioritizes privacy, as no user data is transmitted. The federated learning is steered by the learning protocol. In the synchronous learning protocol, the learning round time is the same for all edge servers, such that all servers can progress synchronously. However, this protocol suffers from the “straggler effect”, where stronger servers will remain idle until weaker servers’ complete one training round. Thus, the asynchronous learning protocol has been introduced as a solution, which enables each server to train at its own pace by relaxing the learning round time. However, this protocol has three drawbacks, as: high communication costs, staleness of some local models, and negative impacts of outdated local models on the global model. To address the drawbacks of both synchronous and asynchronous protocols, the semi-synchronous learning protocol has been proposed. However, existing semi-synchronous protocols are still faced with challenges in fully addressing the above drawbacks. In decentralized topologies, peer-to-peer model exchange allows participating edge servers to continuously update and refine their local models, improving local consistency and reducing dependence on a central coordinator. At the same time, centralized aggregation mechanisms can facilitate global model synchronization and help maintain overall convergence across the network. In this study, a new semi-synchronous federated learning protocol, named SSFL, is presented, which trains a global model by taking all servers into account and incorporating a decentralized topology to balance the accuracy of local models. Additionally, it uses a new client selection mechanism to ensure that the information from all servers is utilized for training the global model. Experiments using the MNIST, Boston Housing, KDD Cup '99, and CIFAR10 datasets demonstrate that the proposed SSFL protocol outperforms existing federated learning protocols in terms of efficient use of all servers’ capacity and information, higher convergence rate of the global model, and lower discrepancy between local models’ accuracy.