依赖差分隐私:关联数据集下的高斯机制

Dependent differential privacy: Gaussian mechanism for correlated datasets

差分隐私(Differential Privacy)是一种数据扰动框架,它保证查询结果在概率上不可区分。研究表明差分隐私应用于关联数据集时,将带来隐私泄露的风险。根据依赖差分隐私(Dependent Differential Privacy),量化了依赖差分隐私敏感度的度量;随后,提出了依赖差分隐私-高斯机制算法(Gaussian Mechanism Algorithm-Dependent Differential Privacy),实现数据扰动,同时证明了该机制满足隐私保证的基本定理;通过使用真实数据集的实验表明,GMA-DDP在管理依赖数据的隐私-效用权衡方面具有较高的可用性。

Differential Privacy is a data perturbation framework,which ensures that the query results are not distinguishable in probability.Research shows that when differential privacy is applied to associated data sets,it will bring the risk of privacy disclosure.Based on the dependent differential privacy,this paper quantifies the sensitivity of the dependent differential privacy;Then,a Gaussian Mechanism Algorithm-Dependent Differential Privacy is proposed to realize data disturbance,and the basic theorem that the mechanism meets the privacy guarantee is proved;Experiments using real data sets show that GMA-DDP has high availability in managing privacy utility tradeoffs that depend on data.