arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2407.20573cs.DC
收藏

面向具有异构模型的分层边缘网络的联邦学习即服务

Federated Learning as a Service for Hierarchical Edge Networks with Heterogeneous Models

Wentao Gao, Omid Tavallaie, Shuaijun Chen, Albert Zomaya

AI总结:

针对IoT设备计算资源异构及非IID数据下异构模型聚合难题,本文提出分层边缘系统异构聚合框架HAF-Edge,通过边缘与云两级聚合及选择性知识转移提升全局模型收敛速度,实验证明其显著优于现有方法。

摘要:

联邦学习(FL)是一种分布式机器学习(ML)框架,能够通过聚合客户端本地训练的模型来训练新的全局模型,而无需共享用户的原始数据。联邦学习即服务(FLaaS)为在具有不同计算资源的设备上训练机器学习模型提供了一种保护隐私的方法。大多数提出的基于FL的方法在所有客户端设备上训练相同的模型,而不考虑它们的计算资源。然而,在实际的物联网(IoT)场景中,计算资源有限的IoT设备可能无法训练由硬件性能更高的客户端设备所承载的模型。大多数旨在解决异构模型聚合问题的现有FL框架都是为独立同分布(IID)数据设计的,当遇到非IID场景时,可能难以达到目标算法性能。为了解决分层网络中的这些问题,本文提出了一种用于分层边缘系统的异构聚合框架,称为HAF-Edge。在我们提出的框架中,我们引入了一种通信高效的模型聚合方法,专为在边缘和云层级运行两级模型聚合的FL系统设计。该方法通过在异构模型聚合期间利用选择性知识转移,增强了全局模型的收敛速度。据我们所知,这项工作开创性地解决了跨越IoT、边缘和云环境的分层FL系统中异构模型聚合的问题。我们进行了大量实验来验证所提出方法的性能。评估结果表明,HAF-Edge显著优于最先进的方法。

英文摘要:

Federated learning (FL) is a distributed Machine Learning (ML) framework that is capable of training a new global model by aggregating clients' locally trained models without sharing users' original data. Federated learning as a service (FLaaS) offers a privacy-preserving approach for training machine learning models on devices with various computational resources. Most proposed FL-based methods train the same model in all client devices regardless of their computational resources. However, in practical Internet of Things (IoT) scenarios, IoT devices with limited computational resources may not be capable of training models that client devices with greater hardware performance hosted. Most of the existing FL frameworks that aim to solve the problem of aggregating heterogeneous models are designed for Independent and Identical Distributed (IID) data, which may make it hard to reach the target algorithm performance when encountering non-IID scenarios. To address these problems in hierarchical networks, in this paper, we propose a heterogeneous aggregation framework for hierarchical edge systems called HAF-Edge. In our proposed framework, we introduce a communication-efficient model aggregation method designed for FL systems with two-level model aggregations running at the edge and cloud levels. This approach enhances the convergence rate of the global model by leveraging selective knowledge transfer during the aggregation of heterogeneous models. To the best of our knowledge, this work is pioneering in addressing the problem of aggregating heterogeneous models within hierarchical FL systems spanning IoT, edge, and cloud environments. We conducted extensive experiments to validate the performance of our proposed method. The evaluation results demonstrate that HAF-Edge significantly outperforms state-of-the-art methods.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑