Research map / Federated learning and privacy / Differential privacy
DP-Fusion: Token-Level Differentially Private Inference for Large Language Models
Rushil Thareja, Preslav Nakov, Praneeth Vepakomma, Nils Lukas
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 |
|---|---|---|---|
| blend private and public next-token distributions with bounded Renyi divergence instead of DP-Prompt | DP-Fusion · TAB-ECHR | perplexity: 8.44 -> 1.426 | Improved |
| use DP-Fusion instead of DP-Decoding | DP-Fusion · TAB-ECHR | perplexity: 3.96 -> 1.426 | Improved |
| relax privacy budget alpha*beta from 0.01 to 0.10 (ablation) | DP-Fusion · TAB-ECHR | perplexity: 1.459 -> 1.426 | Improved |
| relax privacy budget alpha*beta from 0.01 to 0.10 (ablation) | DP-Fusion · TAB-ECHR | LOSS attack ASR: 0.2600 -> 0.2933 | Hurt |
| use DP-Fusion instead of NER scrubbing at strictest privacy | DP-Fusion · TAB-ECHR | LOSS attack ASR: 0.2767 -> 0.2600 | Improved |
| raise alpha*beta from 0.01 to 0.1 (ablation) | DP-Fusion · TAB-ECHR LLM-judge win rate vs public baseline | win rate: 45% -> 56% | Improved |
| use DP-Prompt clipping width 5 instead of 50 (ablation) | DP-Prompt · TAB-ECHR | perplexity: >100 | Hurt |
| raise DP-Prompt temperature from 0.75 to 1.75 (w=50) (ablation) | DP-Prompt · TAB-ECHR | perplexity: 4.26 -> 8.44 | Hurt |
| raise DP-Decoding lambda from 0.1 to 0.9 (ablation) | DP-Decoding · TAB-ECHR | LOSS attack ASR: 0.1567 -> 0.6600 | Hurt |
| label retrieved chunks as privacy groups | DP-Fusion · RAG prompt injection | attack success rate: 90% -> 0% | Improved |
| add more privacy groups m | DP-Fusion · TAB-ECHR | privacy-utility smoothness | Mixed |
Closest papers by meaning
- Split-and-Denoise: Protect large language model inference with local differential privacy (ICML 2024 · similarity 0.91)
- SynBench: A Benchmark for Differentially Private Text Generation (NeurIPS 2026 · similarity 0.90)
- Open LLMs are Necessary for Current Private Adaptations and Outperform their Closed Alternatives (NeurIPS 2024 · similarity 0.89)
- InvisibleInk: High-Utility and Low-Cost Text Generation with Differential Privacy (NeurIPS 2025 · similarity 0.89)
- Beyond Memorization: Violating Privacy via Inference with Large Language Models (ICLR 2024 · similarity 0.89)
- Synthesizing Privacy-Preserving Text Data via Finetuning *without* Finetuning Billion-Scale LLMs (ICML 2025 · similarity 0.89)
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.
