摘要
针对在数据异构和资源异构的无线网络中联邦学习训练效率低及训练能耗高的问题,面向图像识别任务,提出了基于优化引导的异步联邦学习算法AFedGuide。利用较高样本多样性的客户端模型的引导作用,提高单轮聚合有效性。采用基于训练状态的模型增量异步更新机制,提高模型更新实时性以及信息整合能力。设计基于模型差异性的训练决策,修正优化方向。仿真结果显示,相较于对比算法,AFedGuide的训练时长平均减少67.78%,系统能耗平均节省65.49%,客户端的准确率方差平均减少25.5%,说明在客户端数据异构和资源异构的无线网络下,AFedGuide可以在较短的训练时间内以较小的训练能耗完成训练目标,并维持较高的训练公平性和模型适用性。
For the problem of low training efficiency and high training energy consumption of federated learning,in wireless networks with heterogeneous data and resources,Asynchronous Federated Learning Algorithm Based on optimization Guidance(AFedGuide)is proposed for image recognition tasks.The guidance of client model with high sample-diversity is used to improve the effectiveness of single-round aggregation.The asynchronous update mechanism with incremental model and training status is used to improve the timeliness of model update and the integration of model information.The optimization direction is revised by the decision of training based on the difference of model.As the simulation results illustrate,compared with those of the contrast algorithm,the training time of AFedGuide decreases 67.78%averagely,the system energy consumption decreases 65.49%averagely,and the variance of clients accuracy decreases 25.5%averagely.It means that,in a wireless network with heterogeneous data and heterogeneous resources clients,AFedGuide can achieve the training target in a short training time and with less training energy consumption,while maintaining the training fairness and model suitability.
作者
张海波
任俊平
蔡磊
邹灿
ZHANG Haibo;REN Junping;CAI Lei;ZOU Can(School of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;Ubiquitous Sensing and Networking Key Laboratory of Chongqing,Chongqing 400065,China;360 Digital Security Technology Group Co.,Ltd.,Beijing 100015,China)
出处
《电讯技术》
北大核心
2024年第6期979-988,共10页
Telecommunication Engineering
基金
国家自然科学基金资助项目(62271094)
重庆市留创计划创新类资助项目(cx2020059)。
关键词
图像识别
异步联邦学习
数据异构
资源异构
优化引导
image recognition
asynchronous federated learning
heterogeneous data
heterogeneous resources
optimization guidance
