EAGER: Foundations for the Systematic Study of Synthetic Data

由于大量数据的可用性，数据驱动模型在自然语言处理、计算机视觉和机器人等各个领域的进步已经成为可能。然而，对隐私权的担忧一直在增长，导致需要在允许数据共享的同时保护个人数据的策略。合成数据提供了一种潜在的解决方案，它保留了原始数据的统计属性，同时删除了任何个人可识别信息。虽然已经提出了许多生成合成数据的方法，但在这一领域仍然缺乏坚实的理论基础。这个项目弥合了理论与实践之间的差距。该项目的新颖之处在于制订了系统研究综合数据的基本原则，并澄清了技术词汇和有关概念。该项目更广泛的意义和重要性在于它能够使机构阐明、执行、评估和验证其合成数据生成方法所需的约束条件，从而显著加快数据共享。随着隐私的加强和实用性的增强，这一框架将塑造一个隐私得到保护、知识共享、基于人工智能的方法真正为人类的改善而蓬勃发展的世界。该项目概念化了合成数据的哲学考虑，建立了合成数据的属性，开发了“合成”的正式定义，并为合成数据开发了一个全面的评估框架，包括精心策划的数据集、指标和基线。这项研究提高了我们对隐私、效用和合成数据的科学理解。该项目还通过为本科生和研究生的研究和推广活动提供新的安全或隐私项目，培养研究与教育的整合，并作为宝贵的教学工具和进入隐私和安全研究领域的绝佳切入点。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。

Advancements in data-driven models in various fields like natural language processing, computer vision, and robotics have been made possible due to the availability of large amounts of data. However, concerns about privacy rights have been growing, leading to the need for strategies that protect individual data while still allowing data sharing. Synthetic data offers a potential solution by preserving the statistical properties of the original data while removing any personally identifiable information. Although many methods for generating synthetic data have been proposed, there is still a lack of solid theoretical foundations in this area. This project bridges the gap between theory and practice. The project's novelties are in developing the fundamental principles for the systematic study of synthetic data, and in clarifying the technical vocabulary and the associated concepts. The project's broader significance and importance is in its ability to allow institutions to articulate, enforce, evaluate, and validate their required constraints for synthetic data generation methodologies, significantly accelerating data-sharing. With heightened privacy and enhanced utility hand in hand, this framework will shape a world where privacy is safeguarded, knowledge is shared, and AI-based methods truly flourish for the betterment of humanity. The project conceptualizes the philosophical considerations of synthetic data, establishes properties of synthetic data, develops formal definitions of what it means to be “synthetic”, and develops a comprehensive evaluation framework for synthetic data that includes curated datasets, metrics, and baselines. The research improves our scientific understanding of privacy, utility, and synthetic data. The project also cultivates the integration of research and education, by providing new security or privacy projects for undergraduate and graduate research and outreach activities, and serves as an invaluable teaching tool and excellent entry point into the field of privacy and security research.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.