Differentially-Private Synthetic Datasets for Confidentiality Protection of Statistical Data

“你的答案将被保密，仅用于统计目的。”当人们问你一些个人信息时，他们经常向你承诺这一点。你有没有想过他们是如何确保遵守诺言的？他们如何在保护您回答的机密性的同时，与其他人共享部分数据集，或者至少共享研究结果？我的研究项目涉及差异隐私，这是一个相对较新且非常严格的标准，用来衡量保护此类数据的机制所提供的保密保证。我致力于新的方法来创建完全合成的数据集(你可以认为它们是假的数据集)，以满足这一保密标准。这些合成数据集将确保没有人能够识别你或推断你对研究中问题的答案，同时足够接近真实的数据集，从而对统计分析有用。我还将使常用的统计推断方法适应这种合成数据集的特殊性。

"Your answers will be kept confidential and will only be used for statistical purposes." People often promise you this when they ask you for some personal information. Did you ever wonder how they make sure to respect their promise? How can they share parts of the dataset with others, or at least the results of the study, while protecting the confidentiality of your answers? My research program concerns differential privacy, a relatively new and very rigorous criterion to measure the confidentiality guarantee offered by a mechanism design to protect such data. I work on new methods to create completely synthetic datasets (you could think of them as fake datasets) which satisfy this confidentiality criteria. These synthetic datasets will ensure that no one can identify you or deduct your answers to the questions in the study, while being close enough to the real datasets to be useful for statistical analyses. I will also adapt usual statistical inference methods to the particularities of such synthetic datasets.