差分隐私下的权重社交网络隐私保护

Protection of privacy of the weighted social network under differential privacy

由于权重社交网络的复杂性和噪声随机性,传统的隐私保护方法无法兼顾社交网络中的隐私和效用需求。针对此问题,融合直方图统计和非交互式差分隐私查询模型,提出社交网络边权重直方图统计发布方法。该方法将边权重统计直方图作为查询结果,并设计低敏感度的边权重拉普拉斯噪声随机扰动算法,实现社交关系的差分隐私保护。为减少噪声量,引入社区结构熵将社交网络的用户节点划分为若干子社区,提出随机扰动改进算法,以社区为单位划分社交关系并注入拉普拉斯噪声,使各个社区序列满足差分隐私,实现从社区层面保护社交关系。此外,利用一维结构熵的特性,衡量算法对权重社交网络的整体隐私保护程度。理论分析和实验结果表明:所提出的隐私保护算法对节点度识别的保护程度均高于对比算法,能够实现更好的隐私保护效果,同时,在大型社交网络中能够满足差分隐私要求,且保持较高的社交网络数据效用。

Due to randomness of noise and complexity of weighted social networks,traditional privacy protection methods cannot balance both data privacy and utility issues in social networks.This paper addresses these problems,combines histogram statistics and non-interactive differential privacy query model,and proposes a statistical releasing method for the histogram of weighted-edges in social networks.This method regards the statistical histogram of weighted-edges as the query result and designs the low-sensitivity Laplace noise random perturbation algorithm,which realizes the differential privacy protection of social relations.In order to reduce errors,the community structure entropy is introduced to divide the user nodes of the social network into several sub-communities,with the improved stochastic perturbation algorithm proposed.The social relationship is divided by community as a unit and Laplace noise is injected,so that each sequence of community satisfies the differential privacy with the social relationship protected from the community level.In addition,the characteristics of one-dimensional structural entropy are used to measure the overall privacy protection degree of the algorithm with respect to the weighted social network.By theoretical analysis and experimental results,the privacy protection algorithm proposed in this paper has a higher protection degree than the comparison algorithm for node degree identification,which achieves a better privacy protection effect.Meanwhile,it can meet the requirements of differential privacy in large social networks and maintain a high data utility of the weighted social network.