Research map / Federated learning and privacy / Differential privacy
PEARL: Differentially Private and Entropy-Aware Regulated Language Generation
Seongho Joo, Hyukhun Koh, Kyomin Jung
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 |
|---|---|---|---|
| apply PEARL confidence-gap-guided redaction and refill on top of DP decoding (ablation) | exponential-mechanism DP decoding (NoRefill) · LeakRAG-Medical eps=3, LLaMA 3.1 8B | HalluScore: 2.52 -> 1.87 | Improved |
| select refill spans by confidence gap instead of at random (ablation) | PEARL · LeakRAG-Medical eps=3, LLaMA 3.1 8B | HalluScore: 1.97 -> 1.87 | Improved |
| select refill spans by confidence gap instead of at random (ablation) | PEARL · LeakRAG-Financial eps=3, LLaMA 3.1 8B | BERTScore: 52.30 -> 51.15 | No clear effect |
| use confidence-gap-guided rewriting at eps=3 | PEARL · LeakRAG-Medical, LLaMA 3.1 8B | HalluScore: 2.04 (non-private few-shot) -> 1.87 | Improved |
| add refill stage after redaction (Refill with GPT vs Full Redaction) (ablation) | PEARL · LeakRAG-Medical eps=3, LLaMA 3.1 8B | GoldAlign: 2.88 -> 3.06 | Improved |
| add redaction (without refill) to DP decoding (ablation) | NoRefill · LeakRAG-Medical eps=3, LLaMA 3.1 8B | HalluScore: 2.52 -> 2.06 | Improved |
| add PEARL on top of refill/redaction at eps=6 (ablation) | PEARL · LeakRAG-Medical eps=6, LLaMA 3.1 8B | HalluScore: 1.92 -> 1.90 | No clear effect |
| increase number of filtered sentences k (ablation) | PEARL · LeakRAG | HalluScore and leaked PII | Improved |
| use PEARL instead of PICL private decoding | private decoding · LeakRAG-Medical eps=8, LLaMA 3.1 8B | Leaked PII: 3.60 -> 0.57 | Improved |
| use PEARL instead of KSA private decoding | private decoding · LeakRAG-Medical eps=6, LLaMA 3.1 8B | Leaked PII: 0.50 -> 0.53 | No clear effect |
| apply DP noise (exponential mechanism) to RAG decoding | RAG LLM · LeakRAG, LLaMA 3.1 8B | HalluScore: 2.04 (non-private) -> 2.52 (NoRefill, eps=3) | Hurt |
Closest papers by meaning
- Hidden No More: Attacking and Defending Private Third-Party LLM Inference (ICML 2025 · similarity 0.89)
- Can LLMs Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity Theory (ICLR 2024 · similarity 0.89)
- Beyond Memorization: Violating Privacy via Inference with Large Language Models (ICLR 2024 · similarity 0.88)
- InvisibleInk: High-Utility and Low-Cost Text Generation with Differential Privacy (NeurIPS 2025 · similarity 0.88)
- Privacy Auditing of Large Language Models (ICLR 2025 · similarity 0.87)
- Split-and-Denoise: Protect large language model inference with local differential privacy (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.
