CAREER: Extending the Foundations of Privacy-Preserving Machine Learning

尽管现代机器学习算法有许多社会效益，但它对个人隐私构成了真正的威胁。差分隐私是一个数学框架，它使设计具有可证明的输入数据集隐私保证的学习算法成为可能。尽管最近在差分私有（DP）机器学习方面取得了进展，但我们目前对DP学习算法基本特征的理解非常有限。这个职业项目提供了一个多方面的研究计划，在DP机器学习的两个重要领域解决广泛的基本问题：(i)随机优化和（ii）联邦学习。第一个是机器学习中最基本的任务之一，第二个是现代机器学习最有前途的应用之一。本项目旨在：1)理解DP随机优化算法的计算和统计限制；2)建立DP随机非凸优化的综合理论，为开发用于现代机器学习的新DP算法提供坚实的基础；3)为DP联邦学习开发新的高效算法范式，提供有意义和可证明的效用保证。同时考虑到用户数据的演变性质和他们参与协作学习的动机。这项研究的结果有望产生下一代隐私保护学习算法，可以实现广泛的实际应用。这个职业项目包括教育和推广活动，如开发关于优化和差分隐私的新研究生课程，以及组织研讨会以了解现代机器学习中对数据隐私的威胁。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。

Despite their numerous societal benefits, modern machine learning algorithms pose real threats to personal privacy. Differential privacy is a mathematical framework that enables designing learning algorithms with provable privacy guarantees for their input datasets. Despite the recent progress in differentially private (DP) machine learning, our current understanding of the fundamental characteristics of DP learning algorithms is very limited. This career project offers a multifaceted research plan that tackles a broad range of fundamental questions in two important areas of DP machine learning: (i) stochastic optimization and (ii) federated learning. The first is one of the most fundamental tasks in machine learning and the second is one of the most promising applications of modern machine learning. This project aims at: 1) understanding the computational and statistical limits of DP stochastic optimization algorithms, 2) building a comprehensive theory for DP stochastic non-convex optimization, which provides a firm basis for developing new DP algorithms for modern machine learning, and 3) developing new, efficient algorithmic paradigms for DP federated learning that offer meaningful and provable utility guarantees, while taking into account the evolving nature of users’ data and their incentives to participate in collaborative learning. The outcomes of this research are expected to yield the next-generation privacy-preserving learning algorithms that can be implemented for widespread practical use. This career project includes educational and outreach activities such as developing new graduate courses on optimization and differential privacy, and organizing workshops to understand the threats to data privacy in modern machine learning.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.