Research map / Large language models / Retrieval and knowledge
Fine-Grained Privacy Extraction from Retrieval-Augmented Generation Systems by Exploiting Knowledge Asymmetry
Yufei Chen, Yao Wang, Haibin Zhang, Tao Gu
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 q2 (explicit retrieved-content prompt) to q1 (ablation) | adversarial query · HCM | ESR: 55.56% -> 93.55% | Improved |
| add q2 to q1 (ablation) | adversarial query · EE | ESR: 81.30% -> 95.65% | Improved |
| add q2 to q1 (ablation) | adversarial query · NQ | ESR: 52.94% -> 80.00% | Improved |
| use knowledge-asymmetry similarity features plus DNN instead of the best baseline (LLM-based judgment) | privacy sentence identification · HCM | ESR: 65.22% -> 93.55% | Improved |
| use knowledge-asymmetry features plus DNN instead of the best baseline (LLM-based judgment) | privacy sentence identification · NQ (multi-domain) | F1-score: 43.90% -> 84.21% | Improved |
| swap standard LLM used for comparison from LLaMA3.1-8B to Qwen3-8B or GPT-4o (ablation) | attack framework · HCM | ESR: 93.55% -> 85.71% (Qwen3-8B), 88.89% (GPT-4o) | Hurt |
| swap standard LLM used for comparison from LLaMA3.1-8B to GPT-4o (ablation) | attack framework · NQ | ESR: 80.00% -> 85.71% | Improved |
| swap retriever bge-large-en for e5-large-v2 or gte-large (ablation) | attack framework · HCM | ESR: 93.55% -> 94.12% / 94.44% | No clear effect |
| swap retriever bge-large-en for e5-large-v2 (ablation) | attack framework · EE | ESR: 95.65% -> 82.14% | Hurt |
| swap retriever bge-large-en for e5-large-v2 (ablation) | attack framework · NQ | ESR: 80.00% -> 84.62% | Improved |
| use a different RAG LLM (Qwen3-8B or GPT-4o) instead of LLaMA3.1-8B (ablation) | attack framework · NQ | ESR: 80.00% -> 87.50% (Qwen3-8B), 76.73% (GPT-4o) | Mixed |
| vary RAG model size (8B, 70B, 405B) and temperature (0.3, 0.6, 0.9) (ablation) | attack framework · HCM/EE/NQ | ESR | No clear effect |
Closest papers by meaning
- On the Vulnerability of Applying Retrieval-Augmented Generation within Knowledge-Intensive Application Domains (ICML 2025 · similarity 0.88)
- SeCon-RAG: A Two-Stage Semantic Filtering and Conflict-Free Framework for Trustworthy RAG (NeurIPS 2025 · similarity 0.87)
- Beyond Memorization: Violating Privacy via Inference with Large Language Models (ICLR 2024 · similarity 0.87)
- Self-Refining Language Model Anonymizers via Adversarial Distillation (NeurIPS 2025 · similarity 0.87)
- Language Models are Advanced Anonymizers (ICLR 2025 · similarity 0.86)
- Follow My Instruction and Spill the Beans: Scalable Data Extraction from Retrieval-Augmented Generation Systems (ICLR 2025 · similarity 0.86)
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.
