AI Researchers Release Guide for Differentially Private Synthetic Data

Natalia Ponomareva, Zheng Xu, H. Brendan McMahan, Peter Kairouz, Lucas Rosenblatt, and Vincent Cohen-Addad published their work on 2026/07/07 in the Journal of Artificial Intelligence Research (JAIR) under view ID 21111 ¹. Their study is titled “How to DP‑Fy Your Data: A Practical Guide to Generating Synthetic Data With Differential Privacy.”

High‑quality data is vital for unlocking AI’s full potential for end users ¹. Villalobos et al. noted in 2024 that finding new sources of such data is getting harder because most publicly available human‑generated data will soon have been used ¹. Public datasets often fail to represent real‑world user interactions; for example, a research speech dataset of contractors interacting with an AI assistant tends to be more homogeneous, well articulated, and self‑censored than commands issued by typical end users ¹. This mismatch means system creators and end users alike need access to data grounded in authentic usage scenarios.

Direct use of user data carries significant privacy risks, which must be addressed before it can be safely employed. Differential Privacy (DP) is a well established framework for reasoning about and limiting information leakage, and is considered a gold standard for protecting user privacy ¹. The focus of this work—Differentially Private Synthetic data—is synthetic data that preserves overall trends of source data while providing strong privacy guarantees.

Ponomareva et al. provide concrete steps using tools like Google’s Private Aggregation Service (PAS). Their guide demonstrates how practitioners can generate DP synthetic data directly from user interactions while maintaining robust protection against reidentification attacks. However, the authors note that balancing utility—preserving statistical properties useful for training machine learning models—with rigorous privacy protection quantified through epsilon values remains challenging. Achieving a low epsilon value typically requires significantly reducing dataset sizes and increasing added noise, making it difficult to balance these competing goals effectively.

The expert notes elaborate on differential privacy: adding carefully calibrated noise during data generation so an adversary cannot learn individual records from synthetic datasets. They also highlight practical considerations such as selecting appropriate epsilon thresholds and choosing aggregation techniques that minimize utility loss while maximizing privacy. These insights are valuable for developers who need to produce synthetic data that respects users’ confidentiality and retains enough structure to train accurate AI systems.

Overall, Ponomareva and colleagues deliver a detailed, actionable framework for generating differentially private synthetic data tailored to real‑user interaction scenarios. Their work bridges a critical gap between privacy preservation and data utility in AI development, offering system creators concrete methods to protect individuals without sacrificing model performance.