Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Is API Access to LLMs Useful for Generating Private Synthetic Tabular Data?

Marika Swanberg, Ryan McKenna, Edo Roth, Albert Cheu, Peter Kairouz

TL;DR

This work investigates whether API access to foundation models can meaningfully improve differentially private (DP) synthetic tabular data generation. It develops two approaches: (i) adapting Private Evolution to the tabular domain with a workload-aware distance, and (ii) a one-shot API-based method that leverages Gemini-generated public data in offline mode. Across experiments, API-based methods do not consistently surpass strong baselines, highlighting the importance of data domain and robust baselines in DP tabular tasks. The study provides practical insights into when API access helps, and suggests avenues for improvement, such as tabular-focused LLMs and hybrid strategies that combine private evolutions with one-shot public-data methods.

Abstract

Differentially private (DP) synthetic data is a versatile tool for enabling the analysis of private data. Recent advancements in large language models (LLMs) have inspired a number of algorithm techniques for improving DP synthetic data generation. One family of approaches uses DP finetuning on the foundation model weights; however, the model weights for state-of-the-art models may not be public. In this work we propose two DP synthetic tabular data algorithms that only require API access to the foundation model. We adapt the Private Evolution algorithm (Lin et al., 2023; Xie et al., 2024) -- which was designed for image and text data -- to the tabular data domain. In our extension of Private Evolution, we define a query workload-based distance measure, which may be of independent interest. We propose a family of algorithms that use one-shot API access to LLMs, rather than adaptive queries to the LLM. Our findings reveal that API-access to powerful LLMs does not always improve the quality of DP synthetic data compared to established baselines that operate without such access. We provide insights into the underlying reasons and propose improvements to LLMs that could make them more effective for this application.

Is API Access to LLMs Useful for Generating Private Synthetic Tabular Data?

TL;DR

This work investigates whether API access to foundation models can meaningfully improve differentially private (DP) synthetic tabular data generation. It develops two approaches: (i) adapting Private Evolution to the tabular domain with a workload-aware distance, and (ii) a one-shot API-based method that leverages Gemini-generated public data in offline mode. Across experiments, API-based methods do not consistently surpass strong baselines, highlighting the importance of data domain and robust baselines in DP tabular tasks. The study provides practical insights into when API access helps, and suggests avenues for improvement, such as tabular-focused LLMs and hybrid strategies that combine private evolutions with one-shot public-data methods.

Abstract

Differentially private (DP) synthetic data is a versatile tool for enabling the analysis of private data. Recent advancements in large language models (LLMs) have inspired a number of algorithm techniques for improving DP synthetic data generation. One family of approaches uses DP finetuning on the foundation model weights; however, the model weights for state-of-the-art models may not be public. In this work we propose two DP synthetic tabular data algorithms that only require API access to the foundation model. We adapt the Private Evolution algorithm (Lin et al., 2023; Xie et al., 2024) -- which was designed for image and text data -- to the tabular data domain. In our extension of Private Evolution, we define a query workload-based distance measure, which may be of independent interest. We propose a family of algorithms that use one-shot API access to LLMs, rather than adaptive queries to the LLM. Our findings reveal that API-access to powerful LLMs does not always improve the quality of DP synthetic data compared to established baselines that operate without such access. We provide insights into the underlying reasons and propose improvements to LLMs that could make them more effective for this application.

Paper Structure

This paper contains 21 sections, 4 equations, 2 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. The role of data domain.
  3. The importance of appropriate baselines.
  4. Preliminaries
  5. Prior Work
  6. GAN-based methods for DP synthetic data
  7. Marginal-based methods for DP synthetic data
  8. Adapting Private Evolution to Tabular Data
  9. Workload-Aware Distance Function
  10. Experimental results
  11. What we learned
  12. Using Gemini-Generated Records as Public Data
  13. Approach overview
  14. $\mathsf{PMW^{pub}}$ liu2021leveraging
  15. Gemini inference
  16. ...and 6 more sections

Figures (2)

  • Figure 1: Top 1-way marginals on private (outlined) and synthetic (yellow) data. Bottom workload error of synthetic data over time for Private Evolution with $\varepsilon = \infty$.
  • Figure 2: (Left) Workload error for baseline methods for generating tabular synthetic data without use of Gemini. (Right) Workload error for baseline methods and our one-shot methods that use API access to Gemini.

Theorems & Definitions (1)

  • Definition 1: Differential Privacy DworkMNS06dwork2016calibrating