面向半异步协同场景的联邦学习数据共享方法  

Data sharing for semi-asynchronous federated learning

作　　者：胡智尧 于淼[1] 田开元 HU Zhiyao;YU Miao;TIAN Kaiyuan(Academy of Military Sciences,Beijing 100091,China)

机构地区：[1]军事科学院,北京100091

出　　处：《海军工程大学学报》2024年第3期108-112,共5页Journal of Naval University of Engineering

摘　　要：在半异步协同场景中,各单位采用联邦学习数据共享方法协同训练模型时存在频率不均衡的问题,原因是各单位的响应时间存在较大差异。响应时间较短的单位能以较快的训练频率更新模型,故半异步协同场景下难以从训练频率较慢的单位学习知识,导致其性能下降。针对此问题,提出了一种可感知响应时间分布的联邦学习算法,通过对参与训练的各单位进行分组,再利用有向非循环图对各个小组进行调度,以并行或串行的方式执行组内迭代训练,实现各单位训练频率均衡化。实验结果表明:该方法较传统的半异步联邦学习数据共享方法,在训练效率和模型预测性能上均有明显改善。In the semi-asynchronous scenario,there is a problem of unbalanced frequency when each unit uses the federated learning data-sharing method for training because the response time of each unit is quite different.Units with less response time can update the model with faster training frequency,so in this case it is difficult to effectively learn knowledge from units with less response time,which affects its performance.To solve this problem,a federated learning algorithm that senses the distribution of response time was proposed.By grouping the units participating in the training,a directed acyclic graph was used to guide each training group to perform iterative training within the group in a parallel or serial manner,so as to achieve a balanced training frequency.Experimental results show that the above mentioned method significantly improves the training efficiency and model prediction compared with the traditional semi-asynchronous federated learning data-sharing method.

关 键 词：联邦学习 训练频率 半异步 响应时间 有向非循环图 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]