How to DP-fy Your Data: A Practical Guide to Generating Synthetic Data With Differential Privacy

Jul 16, 2026

Upload by Patrick

High-quality, representative real-user data is increasingly difficult to obtain, yet direct use creates substantial privacy risks. Differential Privacy (DP) provides a rigorous framework to bound information leakage. This work centers on differentially private synthetic data, which preserves source-level trends while offering strong individual-level privacy guarantees and enabling safer reuse of data that were previously limited by ad-hoc anonymization. The survey systematizes techniques across modalities (image, tabular, text, federated) and the full system components, from data preparation to privacy tracking and empirical testing.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages102

Reading Time4 h 17 min

Introduction

DP synthetic data and motivation

Survey scope: modalities and techniques

System components and privacy evaluation

arXiv :2512 .03238v2 [ cs .CR] 9 Jul 2026 How to DP-fy Your Data: A Practical Guide to Generating Synthetic Data With Differential Privacy NATALIA PONOMAREVA∗ , Google Research, USA ZHENG XU, Google Research, USA H. BRENDAN MCMAHAN, Google Research, USA PETER KAIROUZ, Google Research, USA LUCAS ROSENBLATT, NYU, USA VINCENT COHEN-ADDAD, Google Research, USA CRISTÓBAL GUZMÁN†, Institute for Mathematical and Computational Engineering, Faculty of Mathematics and School of Engineering, Pontificia Universidad Católica de Chile, Chile RYAN MCKENNA, Google Research, USA GALEN ANDREW, Google Research, USA ALEX BIE, Google Research, USA DA YU, Google Research, USA ALEX KURAKIN, Google DeepMind, USA MORTEZA ZADIMOGHADDAM, Google Research, USA SERGEI VASSILVITSKII, Google Research, USA ANDREAS TERZIS, Google DeepMind, USA High quality data is of vital importance for unlocking the full potential of AI for end users. Villalobos et al. stated in 2024 that finding new sources of such data is getting harder as most publicly-available human generated data will soon have been used. Additionally, publicly available data often is not representative of users of a particular system —for example, a research speech dataset of contractors interacting with an AI assistant will likely be more homogeneous, well articulated and self-censored that real world commands that end users will issue. Therefore unlocking high-quality data grounded in ∗ Corresponding Author. †Work done while at Google ‡Work done at Google as part of student researcher engagement Authors’ Contact Information: Natalia Ponomareva, orcid: 0009-0005-6761-1468, [nponomareva@google.com](nponomareva@google.com), Google Research, USA; Zheng Xu, orcid: 0009-0003-6747-3953, [xuzhustc@gmail.com](xuzhustc@gmail.com), Google Research, USA; H. Brendan McMahan, orcid: 0009-0003-5892-4193, [mcmahan@google.com](mcmahan@google.com), Google Research, USA; Peter Kairouz, orcid: 0000-0001-6897-5937, [kairouz@google.com](kairouz@google.com), Google Research, USA; Lucas Rosenblatt, [lurosenb@google.com](lurosenb@google.com), NYU , New York, NY, USA; Vincent Cohen-Addad, [cohenaddad@google.com](cohenaddad@google.com), Google Research, New York, NY, USA; Cristóbal Guzmán, orcid: 0000-0002-1498-2055, [crguzmanp@uc.cl](crguzmanp@uc.cl), Institute for Mathematical and Computational Engineering, Faculty of Mathematics and School of Engineering, Pontificia Universidad Católica de Chile, Santiago, Chile; Ryan McKenna, orcid: 0000- 0002-4950-1952, [mckennar@google.com](mckennar@google.com), Google Research, USA; Galen Andrew, orcid: 0009-0004-9804-7882, [galenandrew@google.com](galenandrew@google.com), Google Research, USA; Alex Bie, [alexbie@google.com](alexbie@google.com), Google Research, USA; Da Yu, [dayuwork@google.com](dayuwork@google.com), Google Research, USA; Alex Kurakin, orcid: 0009-0008-5952-9394, [kurakin@google.com](kurakin@google.com), Google DeepMind, Mountain View, CA, USA; Morteza Zadimoghaddam, orcid: 0000-0003-0717-1120, [zadim@google.com](zadim@google.com), Google Research, USA; Sergei Vassilvitskii, orcid: 0000-0003-0235-1624, [sergeiv@google.com](sergeiv@google.com), Google Research, USA; Andreas Terzis, orcid: 0000-0002-5681-3399, [aterzis@google.com](aterzis@google.com), Google DeepMind, Mountain View, CA, USA. This work is licensed under a Creative Commons Attribution International 4 .0 License. © 2026 Copyright held by the owner/author(s) . doi: 10.1613/jair.1.21111 Journal of Artificial Intelligence Research, Vol. 86, Article 17 . Publication date: July 2026 . 17:2 • Ponomareva, Xu, McMahan, Kairouz et al. real user interactions is of vital interest to both system creators and end users themselves. However, the direct use of user data comes with significant privacy risks, which must be addressed before the data can be used. Differential Privacy (DP) is a well established framework for reasoning about and limiting information leakage, an

FAQs

Why is synthetic data with differential privacy needed instead of using real user data directly?

Real user interactions are valuable, but direct use creates significant privacy risks. Differentially private synthetic data addresses this by providing strong privacy guarantees while still capturing overall trends from the source dataset.

What does the document mean by “differentially private synthetic data”?

It refers to synthetic data that preserves the overall trends of the source (often user-generated) data, while providing strong privacy guarantees to individuals who contributed to the source dataset.

Which data modalities and generation approaches does the survey cover?

The survey examines techniques and privacy protections for multiple modalities, including image, tabular, text, and federated (decentralized) data, and outlines components needed to build DP synthetic-data systems.

Jul 16, 2026

Upload by Patrick