摘要
针对现有联邦学习同步更新模型不易实现,以及客户端数据分布差异较大时服务器集中融合客户端模型将带来负效应,导致全局模型性能较差的问题,提出了一种基于移动式服务器的联邦学习框架。首先在客户端使用梯度下降法,在服务器则基于模型知识迁移提出移动式联邦融合算法,实现移动异步更新。其次基于移动式服务器的联邦学习框架构建个性化联邦学习机制,解决非独立同分布(non independent and identically distributed,Non-IID)设置时客户端本地模型分类性能较低的问题。最后在3种基准数据集上进行仿真验证,结果表明,与现有方法对比,所提算法实现全局模型分类的精度及所需的通信轮数均优于基线方法。
Aiming at the problem that the existing federated learning synchronous update model is not easy to implement,and when the client data distribution is quite different,the centralized integration of the server and the client model will bring negative effects,resulting in poor performance of the global model,a federated learning framework based on mobile server was proposed.Firstly,the gradient descent method was used in the client,and the mobile federation fusion algorithm was proposed based on model knowledge migration in the server to realize mobile asynchronous update.Secondly,a personalized federated learning mechanism was constructed based on the federated learning framework of mobile server to solve the problem of low classification performance of client local model in Non-IID(non independent and identically distributed)setting.Finally,the simulation results on three benchmark data sets show that compared with the existing methods,the proposed algorithm achieves the global model classification accuracy and the number of communication rounds are better than the baseline method.
作者
吴兰
张亚可
龚利爽
李斌全
WU Lan;ZHANG Yake;GONG Lishuang;LI Binquan(College of Electrical Engineering, Henan University of Technology, Zhengzhou 450001, China)
出处
《中国科技论文》
CAS
北大核心
2022年第3期288-294,共7页
China Sciencepaper
基金
国家自然科学基金资助项目(61973103)
河南省软科学研究计划项目(212400410005)。
关键词
联邦学习
梯度下降
客户端
移动式服务器
个性化联邦学习
federated learning
gradient decrease
client
mobile server
personalized federated learning
