Research map / Federated learning and privacy / Differential privacy
Secret-Protected Evolution for Differentially Private Synthetic Text Generation
Tianze Wang, Zhaoyu Chen, Jian Du, Yingtai Xiao, Linjun Zhang, Qiang Yan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace uniform GDP voting (Aug-PE) with secret clustering and secret protection noise | SecPE, GPT-2 · PubMed random words, r/p=2 | BERT-small accuracy: 24.93 -> 29.18 | Improved |
| tighten privacy r/p from infinity to 2 (ablation) | Aug-PE vs SecPE_2000, GPT-2 · PubMed random words | BERT-small accuracy: Aug-PE 29.70 -> 24.93; SecPE 29.19 -> 29.18 | Mixed |
| use SecPE clustering in the non-private setting | SecPE · PubMed random words, r/p=infinity | BERT-small accuracy: 29.70 -> 29.19 (GPT-2) | Hurt |
| replace per-sample voting by cluster representatives | SecPE · Histogram computation time | seconds per epoch: 30126.4 -> 2.3 (Yelp) | Improved |
| swap generator to Aug-PE vs SecPE on LLM sampling time | SecPE · LLM generation time | seconds per epoch: 1698.7 -> 1693.1 (OpenReview) | No clear effect |
| vary number of clusters K (ablation) | SecPE · PubMed, OpenReview, Yelp | downstream accuracy, FID | No clear effect |
| use SecPE on PII secrets in Yelp | SecPE, GPT-2 · Yelp PII, rating, r/p=10 | accuracy: 65.89 -> 61.22 (rating); 73.60 -> 75.05 (category) | Mixed |
| use stronger generator LLM (GPT-4o-Mini, Qwen-2.5-7B) (ablation) | SecPE_600 · Yelp r/p=10 | rating accuracy: 58.36 -> 62.96 (GPT2 to GPT-4o-Mini); 62.08 -> 63.06 (Qwen 1.5B to 7B) | Improved |
| use 7B Mistral instead of a smaller generator (ablation) | SecPE_600 · Yelp r/p=10 | category/rating accuracy: 72.52, 58.10 vs GPT2 73.82, 58.36 | No clear effect |
| use SecPE instead of Aug-PE at tight budgets on real-synthetic similarity | SecPE · PubMed FID | FID | Improved |
Closest papers by meaning
- SynBench: A Benchmark for Differentially Private Text Generation (NeurIPS 2026 · similarity 0.91)
- InvisibleInk: High-Utility and Low-Cost Text Generation with Differential Privacy (NeurIPS 2025 · similarity 0.89)
- Synthesizing Privacy-Preserving Text Data via Finetuning *without* Finetuning Billion-Scale LLMs (ICML 2025 · similarity 0.89)
- Private Evolution Converges (NeurIPS 2025 · similarity 0.89)
- Differentially Private Synthetic Data via Foundation Model APIs 2: Text (ICML 2024 · similarity 0.88)
- PCEvolve: Private Contrastive Evolution for Synthetic Dataset Generation via Few-Shot Private Data and Generative APIs (ICML 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
