Research map / Federated learning and privacy / Privacy attacks and secure learning
EPSVec: Efficient and Private Synthetic Data Generation via Dataset Vectors
Amin Banayeeanzade, Qingchuan Yang, Deqing Fu, Spencer Hong, Erin Babinsky, Alfy Samuel, Anoop Kumar, Robin Jia, Sai Praneeth Karimireddy
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 |
|---|---|---|---|
| add dataset vector steering to pretrained 2-fixed-shot generation (ablation) | EPSVec++ · IMDb | MAUVE: 28.6 -> 67.8 | Improved |
| add DP-histogram fixed-shot selection (ablation) | EPSVec++ · IMDb | MAUVE: 48.2 -> 67.8 | Improved |
| add DP-histogram selection of fixed shots (vs shots not selected privately) (ablation) | EPSVec++ · IMDb | MAUVE: 42.9 -> 67.8 | Improved |
| use pretrained instead of instruction-tuned model (ablation) | EPSVec · IMDb (eps=3) | MAUVE: 7.8 -> 67.8 | Improved |
| add dataset vector steering to instruction-tuned fixed-shots (ablation) | EPSVec (IT) · IMDb | MAUVE: 0.8 -> 7.8 | Improved |
| add dataset vector steering to zero-shot (ablation) | EPSVec++ w/o fixed shots · IMDb | MAUVE: 19.2 -> 48.2 | Improved |
| add dataset vector steering on Olmo-3-1025-7B | EPSVec++ · Yelp (eps=5) | MAUVE: 72.7 -> 82.7 | Improved |
| add dataset vector steering on Llama-3.1-8B | EPSVec++ · Yelp (eps=5) | MAUVE: 44.1 -> 69.4 | Improved |
| add dataset vector steering on Qwen-3-4B-Base | EPSVec++ · Yelp (eps=5) | MAUVE: 33.1 -> 51.3 | Improved |
| add dataset vector steering on gemma-3-4b-pt | EPSVec++ · Yelp (eps=5) | MAUVE: 27.0 -> 61.7 | Improved |
| use two real shots during vector extraction | EPSVec++ · IMDb (eps=5) | MAUVE: 72.3 -> 64.6 | Hurt |
| use two real shots during vector extraction | EPSVec++ · IMDb (eps=3) | MAUVE: 67.8 -> 70.2 | Improved |
Closest papers by meaning
- SynBench: A Benchmark for Differentially Private Text Generation (NeurIPS 2026 · similarity 0.89)
- Struct-Bench: A Benchmark for Differentially Private Structured Text Generation (NeurIPS 2025 · similarity 0.89)
- PrE-Text: Training Language Models on Private Federated Data in the Age of LLMs (ICML 2024 · similarity 0.89)
- Synthesizing Privacy-Preserving Text Data via Finetuning *without* Finetuning Billion-Scale LLMs (ICML 2025 · similarity 0.89)
- Synthetic Text Generation for Training Large Language Models via Gradient Matching (ICML 2025 · similarity 0.87)
- Differentially Private Synthetic Data via Foundation Model APIs 2: Text (ICML 2024 · similarity 0.87)
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.
