摘要
在联邦网络中,各个设备间数据的数据呈现Non-IID(Non-Independent and Identically Distributed,非独立同分布),同时设备之间的算力状况也不同,这些异构特征会导致各个节点设备学习的模型质量好坏不一,传统联邦学习将各个节点训练结果等权重进行平均融合,使联邦学习效果不理想,针对联邦网络中这些异构特征,论文通过设计实验,将强化学习加入联邦学习的模型融合过程,代替传统的平均融合方法,为融合模型分配动态权值,提高了联邦学习在异构特征环境下的学习效果,增强了联邦学习的鲁棒性。
In a federated network,the data of each device is Non-IID(Non-Independent and Identically Distributed),and the computing power between devices is also different.These heterogeneous features will cause the quality of the model learned by each node device to be good or bad.The traditional federated learning averages the weights of the training results of each node and makes the federated learning effect unsatisfactory.Aiming at these heterogeneous features in the federated network,this article designs experiments to add reinforcement learning to the federated learning.The model fusion process replaces the traditional average fusion method,improves the learning effect of federated learning in a heterogeneous feature environment,and enhances the robustness of federated learning.
作者
禹发
赵潇楚
谷牧
张元杰
YU Fa;ZHAO Xiaochu;GU Mu;ZHANG Yuanjie(College of Computer Science and Technology,China University of Petroleum(East China),Qingdao 266580;Sino-German Institute of Intelligent Technologies,Qingdao 266071;Aerospace Cloud Network Technology Development Co.,Ltd.,Beijing 100039;Qingdao Windaka Technology Co.,Ltd.,Qingdao 266555)
出处
《计算机与数字工程》
2023年第10期2305-2308,2329,共5页
Computer & Digital Engineering
基金
国家自然科学基金项目(编号:62072469)
国家重点研发计划(编号:2018YFE0116700)
山东省自然科学基金项目(编号:ZR2019MF049)
中央高校基本科研业务费专项资金(编号:2015020031)
西海岸人工智能技术创新中心建设专项(编号:2019-1-5,2019-1-6)
上海可信工业控制平台开放项目(编号:TICPSH202003015-ZC)资助。
关键词
异构特征
联邦学习
强化学习
模型融合
深度学习
heterogeneous features
federated learning
reinforcement learning
model fusion
deep learning
