CIF: Small: Collaborative Research: Generative Adversarial Privacy: A Data-driven Approach to Guaranteeing Privacy and Utility

为了公共利益（通过数据驱动的研究）和私人利益（企业数据共享），越来越需要发布数据集。然而，消费者隐私问题在很大程度上阻碍了这种努力，因为大型数据集也包含有关参与个人的机密信息。该项目利用了直接从数据集中学习生成模型的最新进展，引入了一种称为生成对抗隐私（GAP）的新框架。GAP将对抗性学习形式化为希望学习最优隐私机制的私有化者与任何意图学习机密特征的统计对手之间的博弈。这种形式化对于评估数据驱动的方法来对抗具有强大推理能力的对手是至关重要的。该项目将包括与霍尼韦尔实验室的互动，以及与斯坦福大学在电力和智慧城市领域的行业合作伙伴的推广和传播。拓展项目包括让初高中女生在亚利桑那州立大学接触社交网络隐私挑战，以及通过斯坦福大学科学拓展项目办公室对K-12教师进行数据科学培训。该项目将重点关注三个基本问题。前两项确保已公布数据中机密特征的隐私性，并涉及开发：(i)捕获一系列对抗能力的大类损失函数的GAP公式的理论限制；（ii） GAP模型的收敛性保证。第三个问题侧重于通过合成数据集来保证身份隐私，使用生成模型（从训练数据生成合成数据）和统计对手类的组合来了解生成具有效用和隐私保证的合成数据集的有效性。该项目的一个关键要素包括对公开可用的数据集和专有数据进行测试。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。

There is a growing need to publish datasets for both public benefit (via data-driven research) and private gains (enterprise data sharing). However, consumer privacy concerns have largely stymied such efforts since large datasets also contain confidential information about participating individuals. This project leverages recent advancements in learning generative models directly from the datasets to introduce a novel framework called generative adversarial privacy (GAP). GAP formalizes adversarial learning as a game between a privatizer that wishes to learn the optimal privacy mechanism and any statistical adversary intent on learning the confidential features. This formalization is crucial to evaluate data-driven approaches against adversaries with strong inferential capabilities. This project will include interactions with Honeywell Labs as well as outreach and dissemination with Stanford industry partners in the electricity and smart cities sector. Outreach programs include exposing middle- and high-school girls to social network privacy challenges at ASU and K-12 teacher training on data science through the Stanford Office of Science Outreach Program.The project will focus on three foundational problems. The first two ensure privacy of confidential features in the published data and involve developing: (i) theoretical limits of the GAP formulation for a large class of loss functions that capture a range of adversarial capabilities; and (ii) convergence guarantees of the proposed GAP model. The third problem focuses on guaranteeing identity privacy via synthetic datasets using a combination of generative models (to generate synthetic data from training data) and classes of statistical adversaries to understand the efficacy of generating synthetic datasets with both utility and privacy guarantees. A key element of this project involves testing on both publicly available datasets as well as proprietary data.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.