基于深度学习的位置大数据统计发布与隐私保护方法  

Statistics release and privacy protection method of location big data based on deep learning

作　　者：晏燕[1,2] 丛一鸣 Adnan Mahmood 盛权政 YAN Yan;CONG Yiming;Adnan Mahmood;SHENG Quanzheng(School of Computer and Communication,Lanzhou University of Technology,Lanzhou 730050,China;Faculty of Science and Engineering,Macquarie University,NSW 2109,Australia)

机构地区：[1]兰州理工大学计算机与通信学院,甘肃兰州730050 [2]麦考瑞大学科学与工程学院,新南威尔士2109

出　　处：《通信学报》2022年第1期203-216,共14页Journal on Communications

基　　金：国家自然科学基金资助项目(No.61762059,No.61862040)。

摘　　要：针对传统位置大数据统计划分发布结构不合理、划分发布方法效率低下的问题,提出一种基于深度学习的位置大数据统计划分结构预测方法和差分隐私发布方法,以提高位置大数据统计划分发布数据的可用性和执行效率。首先对二维空间进行细致划分和自底向上合并,从而构建合理的空间划分结构。然后将划分结构矩阵组织为三维时空序列,借助深度学习模型提取时空特征,实现对划分发布结构的预测。最后结合预测划分发布结构进行差分隐私预算分配和Laplace噪声添加,实现位置大数据统计划分发布信息的隐私保护。通过实际位置大数据集的实验,证明了所提方法在提高发布数据查询精度和运行效率方面的优势。Aiming at the problems of the unreasonable structure and the low efficiency of the traditional statistical partition and publishing of location big data,a deep learning-based statistical partition structure prediction method and a differential publishing method were proposed to enhance the efficacy of the partition algorithm and improve the availability of the published location big data.Firstly,the two-dimensional space was intelligently partitioned and merged from the bottom to the top to construct a reasonable partition structure.Subsequently,the partition structure matrices were organized as a three-dimensional spatio-temporal sequence,and the spatio-temporal characteristics were extracted via the deep learning model in a bid to realize the prediction of the partition structure.Finally,the differential privacy budget allocation and Laplace noise addition were implemented on the prediction partition structure to realize the privacy protection of the statistical partition and publishing of location big data.Experimental comparison of the real location big data sets proves the advantages of the proposed method in improving the querying accuracy of the published location big data and the execution efficiency of the publishing algorithm.

关 键 词：数据发布隐私保护 位置隐私 隐私空间分解 差分隐私 深度学习 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]