EAGER: Bridging The Gap between Theory and Practice in Data Privacy

该项目旨在弥合隐私保护数据共享和分析方面的理论和实践之间的差距。组织和机构收集的数据是当今信息时代的关键资源。然而，这些数据的泄露对个人隐私构成了严重威胁。虽然差异隐私为开发在数据共享中平衡隐私和效用的技术提供了坚实的基础，但目前在这一领域的研究在理论和实践之间存在着明显的差距。在目前的技术水平下，每项任务都需要专门的算法来实现隐私和效用之间的可接受的权衡。设计新算法的过程是手动的，具有挑战性。此外，这一领域的研究往往采取纯理论方法或纯实验方法；两者都有很大的局限性。该项目旨在开发可广泛和自动应用的算法，以及将理论分析与实验验证相结合的方法，专注于具体(而不是渐近)分析，其中拼写出常量。数据隐私技术的进步将使社会受益，因为它在为公共利益发布数据的需要和保护个人隐私的需要之间提供了更好的平衡。该项目追求以下研究目标，以推进数据隐私的最新技术。一个目标是开发一种通用方法，可以将非私有数据分析算法作为黑盒，并使其私有。这可能需要发展一种比差异隐私更宽松的数据隐私概念。另一个目标是开发一种具体的方法来理解数据分析算法的实用性。由于多种原因，证明渐近效用界的理论方法是有限的。渐近分析忽略了常量(经常也忽略多对数项)，这对实践中的实用是至关重要的。当应用方法需要不可接受的空间和时间计算资源时，除了非常大的参数外，具有吸引人的渐近效用界的方法通常执行得很差。由于所有数据集(包括病态数据集)都必须满足效用界限，因此这样的界限可能非常松散，一旦插入实际参数，它们就没有意义了。弥补这一差距需要更好地了解影响效用的因素、更好的效用指标，以及将效用与数据集特性的依赖关系正式化的方法。所得到的具体方法将理论分析与启发式近似和实验验证相结合，可以更有效地指导实际有效的算法的开发。

This project aims to bridge the gap between theory and practice in privacy-preserving data sharing and analysis. Data collected by organizations and agencies are a key resource in today's information age. However, the disclosure of those data poses serious threats to individual privacy. While differential privacy provides a solid foundation for developing techniques to balance privacy and utility in data sharing, currently there is a significant gap between theory and practice in research in this area. In the current state of the art, each task requires specialized algorithms to achieve acceptable trade-off of privacy and utility. The process of designing new algorithms is manual and challenging. Furthermore, research in this area tends to take either a pure theoretical approach or a pure experimental approach; both have significant limitations. This project aims to develop algorithms that can be broadly and automatically applied, and methodologies for combining theoretical analysis with experimental validations, focusing on concrete (instead of asymptotic) analysis where constants are spelled out. Advances in data privacy techniques will benefit society by providing a better balance between the need to release data to serve public interest and the need to protect individuals' privacy. The project pursues the following research goals to advance the state-of-the-art of data privacy. One goal is to develop a general method that can take a non-private data analysis algorithm as a blackbox, and make it private. This may require the development of a data privacy notion that is more relaxed than differential privacy. Another goal is to develop a concrete approach to understanding the utility of data analysis algorithms. The theoretical approach of proving asymptotic utility bounds is limited for a number of reasons. Asymptotic analysis ignores constants (and oftentimes poly-logarithmic terms as well), which are critical for utility in practice. A method with an appealing asymptotic utility bound often performs poorly except for very large parameters, when applying the method requires an unacceptable amount of space and time computing resources. As the utility bound must hold for all datasets (including pathological ones), such bounds can be so loose that they are meaningless once the actual parameters are plugged in. Bridging this gap requires better understanding of the factors affecting utility, better utility metrics, and methods to formalize the dependencies of utility on dataset features. The resulting concrete approach combines theoretical analysis with heuristic approximations and experimental validations, and can more effectively guide the development of practically effective algorithms.