通信学报 ›› 2018, Vol. 39 ›› Issue (1): 70-77.doi: 10.11959/j.issn.1000-436x.2018013

• 学术论文 • 上一篇    下一篇

MapReduce框架下支持差分隐私保护的随机梯度下降算法

俞艺涵,付钰,吴晓平   

  1. 海军工程大学信息安全系,湖北 武汉 430033
  • 修回日期:2017-12-19 出版日期:2018-01-01 发布日期:2018-02-07
  • 作者简介:俞艺涵(1992-),男,浙江金华人,海军工程大学博士生,主要研究方向为信息系统安全、隐私保护等。|付钰(1982-),女,湖北武汉人,博士,海军工程大学副教授、硕士生导师,主要研究方向为信息安全风险评估等。|吴晓平(1961-),男,山西新绛人,博士,海军工程大学教授、博士生导师,主要研究方向为信息安全、密码学等。
  • 基金资助:
    国家自然科学基金资助项目(61100042);国家社科基金资助项目(15GJ003-201)

Stochastic gradient descent algorithm preserving differential privacy in MapReduce framework

Yihan YU,Yu FU,Xiaoping WU   

  1. Department of Information Security,Naval University of Engineering,Wuhan 430033,China
  • Revised:2017-12-19 Online:2018-01-01 Published:2018-02-07
  • Supported by:
    The National Natural Science Foundation of China(61100042);The National Social Science Foundation of China(15GJ003-201)

摘要:

针对现有分布式计算环境下随机梯度下降算法存在效率性与私密性矛盾的问题,提出一种 MapReduce框架下满足差分隐私的随机梯度下降算法。该算法基于MapReduce框架,将数据随机分配到各个Map节点并启动Map分任务独立并行执行随机梯度下降算法;启动Reduce分任务合并满足更新要求的分目标更新模型,并加入拉普拉斯随机噪声实现差分隐私保护。根据差分隐私保护原理,证明了算法满足ε-差分隐私保护要求。实验表明该算法具有明显的效率优势并有较好的数据可用性。

关键词: 机器学习, 随机梯度下降, MapReduce, 差分隐私保护, 拉普拉斯机制

Abstract:

Aiming at the contradiction between the efficiency and privacy of stochastic gradient descent algorithm in distributed computing environment,a stochastic gradient descent algorithm preserving differential privacy based on MapReduce was proposed.Based on the computing framework of MapReduce,the data were allocated randomly to each Map node and the Map tasks were started independently to execute the stochastic gradient descent algorithm.The Reduce tasks were appointed to update the model when the sub-target update models were meeting the update requirements,and to add Laplace random noise to achieve differential privacy protection.Based on the combinatorial features of differential privacy,the results of the algorithm is proved to be able to fulfill ε-differentially private.The experimental results show that the algorithm has obvious efficiency advantage and good data availability.

Key words: machine learning, stochastic gradient descent, MapReduce, differential privacy preserving, Laplace mechanism

中图分类号: 

No Suggested Reading articles found!