Private Seeds, Public LLMs: Realistic and Privacy-Preserving Synthetic Data Generation
About
Large language models (LLMs) have emerged as a powerful tool for synthetic data generation. A particularly important use case is producing synthetic replicas of private text, which requires carefully balancing privacy and utility. We propose Realistic and Privacy-Preserving Synthetic Data Generation (RPSG), which uses private seeds and integrates privacy-preserving strategies, including a formal differential privacy (DP) mechanism in the candidate selection, to generate realistic synthetic data. Comprehensive experiments against state-of-the-art private synthetic data generation methods demonstrate that RPSG achieves high fidelity to private data while providing strong privacy protection.
Qian Ma, Sarah Rajtmajer• 2026
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Next-token prediction | Pubmed | Next Token Accuracy36.1 | 40 | |
| Next-word prediction | Accuracy35.9 | 12 | ||
| Distributional and Semantic Similarity Evaluation | FID0.07 | 12 | ||
| Membership Inference Attack | Reddit (test) | AUC (PPL)54.3 | 12 | |
| Lexical Diversity | Self-BLEU41 | 12 | ||
| Distributional and Semantic Similarity Evaluation | Pubmed | FID0.03 | 8 | |
| Lexical Diversity | Pubmed | Self-BLEU33 | 8 | |
| Membership Inference Attack | PubMed (test) | AUC (PPL)53.3 | 8 |
Showing 8 of 8 rows