Evaluating Differentially Private Generation of Domain-Specific Text

2025-08-28

AI Summary

In high-stakes domains (e.g., healthcare, finance), generative AI deployment is hindered by privacy regulations and the inability to use real data under strict differential privacy (DP) constraints. Method: This paper introduces the first unified benchmark for DP-compliant text synthesis, integrating five domain-specific text corpora. It systematically evaluates pretraining effects and generation performance under realistic privacy budgets (ε ≤ 1), and establishes a multidimensional evaluation framework covering linguistic quality, task utility, and distributional fidelity. Contribution/Results: Experiments reveal a substantial degradation in both practical utility and distributional fidelity of existing DP text generation methods under stringent privacy constraints—highlighting critical technical bottlenecks. This work establishes the first standardized evaluation framework for DP text synthesis and identifies key directions for improvement, thereby providing both theoretical foundations and empirical benchmarks for jointly optimizing privacy guarantees and data utility.

Technology Category

• Natural Language Processing: Generation
• Machine Learning: Privacy
• Philosophy and Ethics of AI: Privacy & Security

Application Category

• Security and Privacy: Large-scale security measurements
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics

Abstract

Generative AI offers transformative potential for high-stakes domains such as healthcare and finance, yet privacy and regulatory barriers hinder the use of real-world data. To address this, differentially private synthetic data generation has emerged as a promising alternative. In this work, we introduce a unified benchmark to systematically evaluate the utility and fidelity of text datasets generated under formal Differential Privacy (DP) guarantees. Our benchmark addresses key challenges in domain-specific benchmarking, including choice of representative data and realistic privacy budgets, accounting for pre-training and a variety of evaluation metrics. We assess state-of-the-art privacy-preserving generation methods across five domain-specific datasets, revealing significant utility and fidelity degradation compared to real data, especially under strict privacy constraints. These findings underscore the limitations of current approaches, outline the need for advanced privacy-preserving data sharing methods and set a precedent regarding their evaluation in realistic scenarios.

Problem

• Evaluating utility and fidelity of differentially private text generation
• Addressing privacy barriers in domain-specific AI data usage
• Benchmarking privacy-preserving methods under realistic constraints

Innovation

• Differentially private synthetic data generation
• Unified benchmark for utility and fidelity
• Evaluation across five domain-specific datasets