Research map / Federated learning and privacy / Differential privacy
SeqPATE: Differentially Private Text Generation via Knowledge Distillation
Zhiliang Tian, Yingxiu Zhao, Ziyue Huang, Yu-Xiang Wang, Nevin Zhang, He He
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 |
|---|---|---|---|
| aggregate teachers by averaging output distributions instead of voting (ablation) | SeqPATE · AirDialog | Bleu-4: 1.85 -> 3.24 | Improved |
| train student with KL divergence instead of cross-entropy on top-1 (ablation) | SeqPATE · AirDialog | Bleu-4: 1.81 -> 3.24 | Improved |
| add pseudo-label loss L_pseudo (ablation) | SeqPATE · AirDialog | Bleu-4: 3.17 -> 3.24 | No clear effect |
| add efficient knowledge distillation (select samples to query teachers) (ablation) | SeqPATE · AirDialog | Bleu-4: 3.02 -> 3.24 | Improved |
| add efficient knowledge distillation (ablation) | SeqPATE · Europarl_v6 | Bleu-4: 0.57 -> 0.61 | Improved |
| Gaussian mechanism instead of Laplace mechanism (ablation) | SeqPATE · AirDialog | Bleu-4: 2.78 -> 3.24 | Improved |
| use all proposed strategies instead of none (original PATE-like) (ablation) | SeqPATE · AirDialog | PPL: 13.21 -> 8.00 | Improved |
| use all proposed strategies instead of none (ablation) | SeqPATE · Europarl_v6 | Bleu-4: 0.44 -> 0.61 | Improved |
| top-p (0.95) filtering instead of top-k=1 (ablation) | SeqPATE · AirDialog | Bleu-4: 0.38 -> 3.24 | Improved |
| top-p filtering instead of top-k=10 (ablation) | SeqPATE · AirDialog | Bleu-4: 1.95 -> 3.24 | Improved |
| top-k=50 instead of top-p filtering (ablation) | SeqPATE · AirDialog | Bleu-4: 3.24 -> 3.35 | No clear effect |
| top-k=200 instead of top-p filtering (oversize k adds noise) (ablation) | SeqPATE · AirDialog | Bleu-4: 3.24 -> 2.94 | Hurt |
Closest papers by meaning
- SynBench: A Benchmark for Differentially Private Text Generation (NeurIPS 2026 · similarity 0.90)
- InvisibleInk: High-Utility and Low-Cost Text Generation with Differential Privacy (NeurIPS 2025 · similarity 0.90)
- Synthesizing Privacy-Preserving Text Data via Finetuning *without* Finetuning Billion-Scale LLMs (ICML 2025 · similarity 0.88)
- Privacy-Preserving In-Context Learning with Differentially Private Few-Shot Generation (ICLR 2024 · similarity 0.87)
- PrE-Text: Training Language Models on Private Federated Data in the Age of LLMs (ICML 2024 · similarity 0.87)
- RAPID: Retrieval Augmented Training of Differentially Private Diffusion Models (ICLR 2025 · 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.
