English

应用科学学报 ›› 2022, Vol. 40 ›› Issue (3): 493-510.doi: 10.3969/j.issn.0255-8297.2022.03.012

• 计算机科学与应用 • 上一篇    下一篇

基于深度学习的联邦集成算法

罗长银1,2,3, 陈学斌1,2,3, 宋尚文1,2,3, 张淑芬1,2,3, 刘之瑜1,2,3   

  1. 1. 华北理工大学 理学院, 河北 唐山 063210;
    2. 华北理工大学 河北省数据科学与应用重点实验室, 河北 唐山 063210;
    3. 华北理工大学 唐山市数据科学重点实验室, 河北 唐山 063210
  • 收稿日期:2020-06-23 发布日期:2022-05-25
  • 通信作者: 陈学斌,博士,教授,研究方向为数据安全、网络安全。E-mail:chxb@qq.com E-mail:chxb@qq.com
  • 基金资助:
    国家自然科学基金(No.61572170,No.61170254,No.61379116)资助

Federated Ensemble Algorithm Based on Deep Neural Network

LUO Changyin1,2,3, CHEN Xuebin1,2,3, SONG Shangwen1,2,3, ZHANG Shufen1,2,3, LIU Zhiyu1,2,3   

  1. 1. School of Science, North China University of Science and Technology, Tangshan 063210, Hebei, China;
    2. Hebei Province Key Laboratory of Data Science and Application, North China University of Science and Technology, Tangshan 063210, Hebei, China;
    3. Tangshan Key Laboratory of Data Science, North China University of Science and Technology, Tangshan 063210, Hebei, China
  • Received:2020-06-23 Published:2022-05-25

PDF

249

摘要： 联邦学习是多源隐私数据保护领域研究的热点，其框架在满足数据不出本地的情况下，可以训练出多方均满意的共同模型，但存在本地模型参数难以整合且无法在安全的情况下将多源数据充分使用的问题，因此提出基于深度学习的联邦集成算法，将深度学习与集成学习应用到联邦学习的框架下，通过优化本地模型的参数，提高了本地模型准确率；使用不同的集成算法来整合本地模型参数，在提升模型准确率的同时兼顾了多源数据的安全性。实验结果表明：与传统多源数据处理技术相比，该算法在mnist、digits、letter、wine数据集训练模型的准确率依次提升1%、8%、-1%、1%，在保证准确率的同时也提升多源数据与模型的安全性，具有很重要的应用价值。

关键词: 联邦学习, 联邦集成算法, 深度神经网络模型, 集成算法, 深度学习

Abstract: Federated learning is a research hotspot in the field of multi-source privacy data protection. It has advantage that its framework can train a common model that is satisfactory to many parties when the data is not local, but it hardly integrates local model parameters and cannot make full use of multiple sources under safe conditions. Aiming to the problem, this article proposes a deep learning-based federated integration algorithm,which applies deep learning and integrated learning to the framework of federated learning. Under the framework of the proposed federated learning, the parameters of local modes are optimized, accordingly the accuracy of the local model is improved. Besides, by applying variety of integration algorithms to integrate local model parameters, the improvement of model accuracy with taking into account the security of multi-source data simultaneously can be achieved. Experimental comparisons with traditional multi-source data processing technology are demonstrated, and show that the accuracies of the proposed training model on the mnist, digits, letter, and wine data sets are increased by 1%, 8%, -1%, 1%, respectively. So that the proposed algorithm could guarantee the improvement of accuracy and the security of multi-source data and models at the same time, and this has important practical application value.

Key words: federated learning, federated ensemble algorithm, deep neural network (DNN) model, ensemble algorithm, deep learning

中图分类号: