Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

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

TaskDatasetResultRank
Next-token predictionPubmed
Next Token Accuracy36.1
40
Next-word predictionREDDIT
Accuracy35.9
12
Distributional and Semantic Similarity EvaluationREDDIT
FID0.07
12
Membership Inference AttackReddit (test)
AUC (PPL)54.3
12
Lexical DiversityREDDIT
Self-BLEU41
12
Distributional and Semantic Similarity EvaluationPubmed
FID0.03
8
Lexical DiversityPubmed
Self-BLEU33
8
Membership Inference AttackPubMed (test)
AUC (PPL)53.3
8
Showing 8 of 8 rows

Other info

Follow for update