Differentially Private Tabular Data Synthesis using Large Language Models
Toan V. Tran, Li Xiong
TL;DR
This work tackles synthetic tabular data generation under differential privacy by introducing DP-LLMTGen, a framework that fine-tunes pretrained large language models in two stages to capture format constraints and data distributions while honoring DP. It employs tabular-to-text encoding, a DP-aware two-stage loss combining Weighted Cross Entropy and Numerical-Understanding losses, and sampling to generate tabular data, with decoding back to table format. Empirical results across five datasets show DP-LLMTGen generally outperforms margins-based and GAN-based baselines in statistical fidelity and downstream ML tasks under $(\epsilon,\delta)$-DP, and the ablation studies confirm the necessity of the two-stage training, the benefit of WCEL and NUL, and the value of feature naming and non-dialogue optimization. Additionally, the framework supports controllable generation to reduce biases in a fairness-aware setting, achieving substantial improvements in demographic parity with minimal utility loss. While effective, the approach incurs high computational cost, motivating future work on efficiency and broader model support.
Abstract
Synthetic tabular data generation with differential privacy is a crucial problem to enable data sharing with formal privacy. Despite a rich history of methodological research and development, developing differentially private tabular data generators that can provide realistic synthetic datasets remains challenging. This paper introduces DP-LLMTGen -- a novel framework for differentially private tabular data synthesis that leverages pretrained large language models (LLMs). DP-LLMTGen models sensitive datasets using a two-stage fine-tuning procedure with a novel loss function specifically designed for tabular data. Subsequently, it generates synthetic data through sampling the fine-tuned LLMs. Our empirical evaluation demonstrates that DP-LLMTGen outperforms a variety of existing mechanisms across multiple datasets and privacy settings. Additionally, we conduct an ablation study and several experimental analyses to deepen our understanding of LLMs in addressing this important problem. Finally, we highlight the controllable generation ability of DP-LLMTGen through a fairness-constrained generation setting.