删除或更新信息,请邮件至freekaoyan#163.com(#换成@)

面向物联网隐私数据分析的分布式弹性网络回归学习算法

本站小编 Free考研考试/2022-01-03

方维维1,,,
刘梦然1,
王云鹏1,
李阳阳2,
安竹林3
1.北京交通大学计算机与信息技术学院 北京 100044
2.社会安全风险感知与防控大数据应用国家工程实验室 北京 100041
3.中国科学院计算技术研究所 北京 100190
基金项目:北京市自然科学基金(L191019),赛尔网络下一代互联网创新项目(NGII20190308)

详细信息
作者简介:方维维:男,1981年生,博士,副教授,研究方向为物联网、边缘计算、分布式机器学习
刘梦然:女,1994年生,硕士生,研究方向为物联网、边缘计算、ADMM
王云鹏:男,1996年生,硕士生,研究方向为物联网、边缘计算、ADMM
李阳阳:男,1987年生,博士,高级工程师,研究方向为移动网络、边缘计算、系统安全
安竹林:男,1980年生,博士,高级工程师,研究方向为物联网、边缘计算、分布式系统
通讯作者:方维维 wwfang@bjtu.edu.cn
中图分类号:TN915; P391.4

计量

文章访问数:1655
HTML全文浏览量:853
PDF下载量:122
被引次数:0
出版历程

收稿日期:2019-09-25
修回日期:2020-05-12
网络出版日期:2020-05-17
刊出日期:2020-10-13

A Distributed Elastic Net Regression Algorithm for Private Data Analytics in Internet of Things

Weiwei FANG1,,,
Mengran LIU1,
Yunpeng WANG1,
Yangyang LI2,
Zhulin AN3
1. School of Computer and Information Technology, Beijing Jiaotong University, Beijing 100044, China
2. National Engineering Laboratory for Public Safety Risk Perception and Control, Beijing 100041, China
3. Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190, China
Funds:Beijing Municipal Natural Science Foundation (L191019), The CERNET Innovation Project (NGII20190308)


摘要
摘要:为了解决基于集中式算法的传统物联网数据分析处理方式易引发网络带宽压力过大、延迟过高以及数据隐私安全等问题,该文针对弹性网络回归这一典型的线性回归模型,提出一种面向物联网(IoT)的分布式学习算法。该算法基于交替方向乘子法(ADMM),将弹性网络回归目标优化问题分解为多个能够由物联网节点利用本地数据进行独立求解的子问题。不同于传统的集中式算法,该算法并不要求物联网节点将隐私数据上传至服务器进行训练,而仅仅传递本地训练的中间参数,再由服务器进行简单整合,以这样的协作方式经过多轮迭代获得最终结果。基于两个典型数据集的实验结果表明:该算法能够在几十轮迭代内快速收敛到最优解。相比于由单个节点独立训练模型的本地化算法,该算法提高了模型结果的有效性和准确性;相比于集中式算法,该算法在确保计算准确性和可扩展性的同时,可有效地保护个体隐私数据的安全性。
关键词:物联网/
隐私保护/
弹性网络回归/
分布式算法/
交替方向乘子法
Abstract:In order to solve the problems caused by the traditional data analysis based on the centralized algorithm in the IoT, such as excessive bandwidth occupation, high communication latency and data privacy leakage, considering the typical linear regression model of elastic net regression, a distributed learning algorithm for Internet of Things (IoT) is proposed in this paper. This algorithm is based on the the Alternating Direction Method of Multipliers (ADMM) framework. It decomposes the objective problem of elastic net regression into several sub-problems that can be solved independently by each IoT node using its local data. Different from traditional centralized algorithms, the proposed algorithm does not require the IoT node to upload its private data to the server for training, but rather the locally trained intermediate parameters to the server for aggregation. In such a collaborative manner, the server can finally obtain the objective model after several iterations. The experimental results on two typical datasets indicate that the proposed algorithm can quickly converge to the optimal solution within dozens of iterations. As compared to the localized algorithm in which each node trains the model solely based on its own local data, the proposed algorithm improves the validity and the accuracy of training models; as compared to the centralized algorithm, the proposed algorithm can guarantee the accuracy and the scalability of model training, and well protect the individual private data from leakage.
Key words:Internet of Things(IoT)/
Privacy protection/
Elastic net regression/
Distributed algorithm/
Alternating Direction Method of Multipliers (ADMM)



PDF全文下载地址:

https://jeit.ac.cn/article/exportPdf?id=50a98a03-22a0-49a9-8e5b-ec748245bf7a
相关话题/物联网 网络 计算 数据 北京