面向数据流发布的数据自适应隐私保护机制  

Data-adaptive privacy-preserving mechanism for data stream publishing in real-time

作　　者：王腾[1] 杨新宇[1] 任雪斌[1] 赵俊 Teng WANG;Xinyu YANG;Xuebin REN;Jun ZHAO(School of Computer Science and Technology,XVan Jiaotong University,Xi'an 710049,China;School of Computer Science and Engineering,Nanyang Technological University,Singapore 639798,Singapore)

机构地区：[1]西安交通大学计算机科学与技术学院,中国西安710049 [2]School of Computer Science and Engineering,Nanyang Technological University,Singapore 639798,Singapore

出　　处：《中国科学：信息科学》2021年第7期1199-1216,共18页Scientia Sinica(Informationis)

基　　金：国家自然科学基金(批准号:61772410,61802298,U1811461);中国博士后基金(批准号:2017M623177);中央高校基本科研业务费(批准号:xjj2018237)资助项目。

摘　　要：群智感知系统中针对数据流的实时发布和深度学习在极大方便人们日常生活的同时,也严重威胁了参与用户的隐私信息.现有隐私保护机制在处理动态性强、时空相关性复杂的数据流时,大都难以实现数据自适应性,从而导致较低的数据效用性.因此,基于ω-事件级差分隐私,本文提出了一种数据自适应的多维数据流隐私保护实时发布机制AdaPub.该机制通过集成基于多重哈希的维度划分策略和自适应累积回溯时间聚类策略分别学习数据流的空间和时间相关性,不需要预定义任何参数,能够根据数据流的动态变化趋势来自适应地调整隐私参数,从而保证了隐私保护机制的数据自适应性并有效提高了数据效用性.此外,本文进一步提出了一种面向层次数据流发布的隐私保护机制HierAdaPub,利用最优隐私预算分配策略来最小化扰动方差以保证数据效用性.大量仿真实验从不同角度均验证了所提出隐私保护机制能够在提供强隐私保护的同时,具有较高的数据效用性.The real-time publishing and deep exploitation of data streams in crowdsensing systems have significantly facilitated people’s daily lives.However,it also seriously compromises the private information of participating users.The existing approaches are non-adaptive to dynamic changes of streams,thus are vulnerable to low data utility.To address such concerns,in this paper,we present AdaPub,a data-adaptive mechanism for infinite multi-dimensional stream real-time publishing underω-event differential privacy.No longer predefining parameters,AdaPub seamlessly incorporates two modules DimParti and AdaCluster to learn spatial and temporal correlations simultaneously in a data-adaptive manner,thus ensuring data adaptability of privacy-preserving mechanism and greatly improving the data utility of the sanitized streams.Moreover,for hierarchical aggregated streams publishing,we further propose a data-adaptive mechanism HierAdaPub that leverages an optimal privacy budget allocation strategy to minimize the total perturbation errors.Extensive experiments on real-world and synthetic datasets demonstrate that our mechanisms substantially outperform the state-of-the-art solutions in terms of both data utility and data adaptivity while achieving strong privacy guarantees.

关 键 词：数据流发布 数据自适应 差分隐私 时空相关性 数据效用性 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]