A Study on Differential Privacy for Data Base with Correlation

差异隐私是将噪声添加到给定查询的答案中的技术。然而，具有相关数据记录的数据库的差异隐私问题并没有得到太多的关注。在这项研究中，我们发现了一个反直觉的现象，即拥有少量相关背景知识的对手比拥有大量相关知识的对手可以获得更多的隐私信息。然后，我们构建了贝叶斯隐私模型来解释这种现象，并对这种情况进行了改进。给出了拉普拉斯分布在差分保密中的适当参数值的近似算法，使得信息泄露的可证明性小于预先确定的阈值；此外，我们还研究了加噪声在差分保密中引起的问题。

The differential privacy is the technology of adding noise to an answer for the given query. The differential privacy for data base with correlated data records, however, has not get much attention. In this research, we find the counter intuitive phenomena that an adversary with small amount of background knowledge about correlations can get more privacy information than the adversary with big amount of these knowledge. Then, we build the Bayesian privacy model which explains this kind of phenomena and improve this situation. We also show the approximation algorithm that gives us the proper parameter value of Laplace distribution employed in differential privacy which make the provability of information leakage less than previously determined threshold.In addition, we investigated the problems which are caused by added noise in differential privacy.