Simulation-Based Inference for Differential Privacy

该研究项目将提供工具，从受隐私限制的数据中获得准确而广泛的统计结论。差别隐私是一种越来越多地被采用的技术，用于保护政府和行业内的数据，例如美国2020年十年一次的人口普查。然而，隐私保护是以对这些数据进行分析的准确性为代价的，有时会严重影响由此产生的决策和结论。在雇用和培训来自不同背景的研究生的同时，该项目将使用计算机模拟技术来处理这些隐私设置下的广泛统计任务。这些新工具提高的准确性将允许更广泛地采用差异隐私，并增加共享数据的可能性，同时降低侵犯隐私的风险。这将保证决策机构以及社会科学和其他学术研究领域的研究人员更广泛地获得基本和可靠的信息。研究结果将通过统计和计算机科学领域的期刊和会议论文集的一系列出版物，以及通过在国家和国际科学会议和讲习班上的发言来传播。开源软件包将被开发并提供给更广泛的社区。该研究项目将提供理论和实践工具，用于复杂参数设置中的统计方法的进步，例如由差异隐私机制的附加噪声所带来的统计方法。差分隐私通过在数据中引入校准噪声（随机性）来保护数据中包含的个人隐私信息。这种机制背后的想法是，即使是消息灵通的攻击者/黑客也无法检测到输出的变化是由于特定个体的反应还是仅仅由于随机性。然而，这些噪声添加技术也会给研究人员的分析带来额外的偏差和方差，这些研究人员想要使用这些数据来提高政府、工业和学术界的知识。该项目将依靠基于模拟的统计方法，如共足抽样和间接推断，提供更准确的分析技术。在保留相同级别的隐私的同时，这种方法将考虑到用于将数据私有化的噪声机制。待开发的工具将通过纠正估计器的偏差和为广泛的统计方法提供可靠的置信区间和假设检验来改进对嘈杂的私有化数据的估计和统计推断。该项目将在统计隐私和基于模拟的推断技术之间建立一些初步联系，并将扩展健壮统计领域。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This research project will deliver tools to obtain accurate and broad statistical conclusions from data that are subject to privacy constraints. Differential Privacy is an increasingly adopted technique to protect data within government and industry, such as in the US 2020 Decennial Census. However, privacy protection comes at a cost in terms of accuracy of the analysis run on these data, sometimes drastically affecting the decisions and conclusions that entail. While employing and training graduate students from diverse backgrounds, this project will use computer-simulation techniques to tackle a wide range of statistical tasks under these privacy settings. The increased accuracy from these new tools will allow for the wider adoption of Differential Privacy and increase the possibility of sharing data with reduced risks of privacy violations. This will guarantee broader access to essential and reliable information for decision-making bodies as well as for researchers in the social sciences and other fields of academic research. Results will be disseminated through a series of publications in journals and conference proceedings in the fields of statistics and computer science, as well as through presentations at national and international scientific conferences and workshops. Open-source software packages will be developed and made available to the broader community.This research project will deliver both theoretical and practical tools for the advancement of statistical approaches in complex parametric settings such as those entailed by the added noise of Differential Privacy mechanisms. Differential Privacy protects the private information of individuals included in the data by introducing calibrated noise (randomness) into the data. The idea behind this mechanism is that even a highly informed attacker/hacker will not be able to detect whether changes in outputs are due to a particular individual's response or are simply due to randomness. However, these noise-addition techniques also introduce additional bias and variance into the analyses made by researchers who will want to use these data for the advancement of knowledge in government, industry, and academia. This project will deliver more accurate analytical techniques by relying on simulation-based statistical methods, such as co-sufficient sampling and indirect inference. While preserving the same level of privacy, this approach will take into account the noise mechanisms used to privatize the data. The tools to be developed will improve estimation and statistical inference on noisy privatized data by correcting bias of estimators and delivering reliable confidence intervals and hypothesis tests for a wide range of statistical methods. The project will establish some of the first links between statistical privacy and simulation-based inference techniques and will expand the field of robust statistics.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.