CAREER: New Frontiers of Private Learning and Synthetic Data

大量收集详细的个人数据为研究人员、公司和政策制定者提供了巨大的好处。为保障个人私隐，许多公营及私营机构均采用差别私隐作为严格的私隐措施。然而，最近对差异隐私的部署揭示了关键的研究空白。首先，现有的差分隐私理论工作大多侧重于最坏情况分析，这往往导致过于悲观的结果，无法为实践中的算法设计提供信息。尽管最近取得了进展，但用于机器学习和数据共享的差异私有算法仍然不是广泛采用的技术。最后，缺乏全面的隐私风险评估工具，使得从业者难以评估差异隐私的有效性并确定适当的隐私风险参数。本项目旨在通过扩展隐私保护算法的曲目和开发审计机制来评估这些算法提供的隐私保护，从而解决差异隐私方面的这些挑战。该研究主要关注两个基本且密切相关的问题：私有学习和私有合成数据。在私有学习中，目标是使用具有差分隐私保证的敏感数据学习准确的机器学习模型。在私有合成数据中，目标是区别地私有地生成一个合成数据集，该数据集保留敏感数据集的重要统计趋势。该项目通过三个研究重点推进了这两个问题的前沿。第一个推力发展了一个理论框架，超越了悲观的最坏情况分析，以更好地捕捉实际场景并指导算法设计。第二个重点是设计实用的算法，这些算法由实践中问题的理论原理和经验结构提供信息。第三部分侧重于隐私攻击和评估学习和合成数据算法的隐私风险的审计机制。该项目还包括一个全面的教育和推广计划，为不同教育水平的学生提供研究机会，并开发新的课程和教材。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。

The vast collection of detailed personal data offers significant benefits to researchers, companies, and policymakers. To protect individual privacy, many organizations, both from the public and private sectors, have adopted differential privacy as a rigorous privacy measure. However, recent deployments of differential privacy have revealed key research gaps. First, much of the existing theoretical work in differential privacy focuses on worst-case analyses, which often lead to overly pessimistic results and fail to inform algorithm design in practice. Despite recent advancements, differentially private algorithms for machine learning and data sharing are still not widely adopted technologies. Lastly, the lack of comprehensive tools for privacy risk assessment makes it difficult for practitioners to evaluate the effectiveness of differential privacy and to determine appropriate privacy risk parameters. This project aims to address these challenges in differential privacy by expanding the repertoire of privacy-preserving algorithms and developing auditing mechanisms to assess the privacy protection these algorithms provide.The research focuses on two fundamental and closely related problems: private learning and private synthetic data. In private learning, the goal is to learn accurate machine learning models using sensitive data with differential privacy guarantees. In private synthetic data, the goal is to differentially privately generate a synthetic dataset that preserves important statistical trends of the sensitive dataset. The project advances the frontiers of these two problems with three research thrusts. The first thrust develops a theoretical framework that goes beyond pessimistic worst-case analyses to better capture practical scenarios and guide algorithm design. The second thrust designs practical algorithms that are informed by theoretical principles and empirical structures of the problems in practice. The third focuses on privacy attacks and auditing mechanisms that evaluate the privacy risks of learning and synthetic data algorithms. The project also includes a comprehensive educational and outreach program, providing research opportunities for students at different educational levels and developing new courses and educational materials.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.