Research map / Large language models / Language, safety and interpretability
ProPILE: Probing Privacy Leakage in Large Language Models
Siwon Kim, Sangdoo Yun, Hwaran Lee, Martin Gubri, Sungroh Yoon, Seong Joon Oh
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 |
|---|---|---|---|
| Tune and prepend a soft prompt using 128 examples instead of relying on five handcrafted templates | OPT-1.3B ProPILE phone-number probing · Pile phone-number evaluation set | exact match rate: 0.0047% -> 1.3% | Improved |
| Increase soft-prompt training examples from 16 to 128 (ablation) | OPT-1.3B ProPILE phone-number probing · Pile phone-number evaluation set | exact match rate: 0.12% to 1.50% | Improved |
| Transfer the OPT-1.3B soft prompt to OPT-350M | OPT-350M ProPILE phone-number probing · Pile phone-number evaluation set | average likelihood: 1.05\!\times\! 10^{-11} -> 1.08\!\times\!10^{-10} | Improved |
| Transfer the OPT-1.3B soft prompt to OPT-1.3B | OPT-1.3B ProPILE phone-number probing · Pile phone-number evaluation set | average likelihood: 6.06\!\times\!10^{-8} -> 3.47\!\times\!10^{-6} | Improved |
| Transfer the OPT-1.3B soft prompt to OPT-2.7B | OPT-2.7B ProPILE phone-number probing · Pile phone-number evaluation set | average likelihood: 1.39\!\times\!10^{-7} -> 2.18\!\times\!10^{-6} | Improved |
| Transfer the OPT-1.3B soft prompt to OPT-1.3B | OPT-1.3B ProPILE phone-number probing · Pile phone-number evaluation set | exact match count: 5 -> 3 | Hurt |
| Transfer the OPT-1.3B soft prompt to OPT-2.7B | OPT-2.7B ProPILE phone-number probing · Pile phone-number evaluation set | exact match count: 14 -> 15 | Improved |
| Use a PII-type word embedding to initialize the soft prompt rather than uniform or mean-vocabulary initialization (ablation) | OPT-1.3B ProPILE phone-number probing · Pile phone-number evaluation set | exact match rate | Improved |
| Increase the number of soft tokens (ablation) | OPT-1.3B ProPILE phone-number probing · Pile phone-number evaluation set | leakage | Improved |
| Provide an additional associated PII item in the prompt, moving from a twin to a triplet (ablation) | OPT-1.3B ProPILE structured-PII probing · Pile email-address and phone-number evaluation sets | exact match rate: fivefold increase for email; more than double for phone numbers | Improved |
| Increase the number of prompt templates (ablation) | OPT-1.3B ProPILE black-box probing · Pile PII evaluation set | exact match rate | Improved |
| Increase beam-search size (ablation) | OPT ProPILE black-box probing · Pile PII evaluation set | target PII generation frequency | Improved |
Closest papers by meaning
- PrivAuditor: Benchmarking Data Protection Vulnerabilities in LLM Adaptation Techniques (NeurIPS 2024 · similarity 0.87)
- Can LLMs Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity Theory (ICLR 2024 · similarity 0.87)
- Beyond Memorization: Violating Privacy via Inference with Large Language Models (ICLR 2024 · similarity 0.87)
- The Canary’s Echo: Auditing Privacy Risks of LLM-Generated Synthetic Text (ICML 2025 · similarity 0.86)
- A Synthetic Dataset for Personal Attribute Inference (NeurIPS 2024 · similarity 0.86)
- Open LLMs are Necessary for Current Private Adaptations and Outperform their Closed Alternatives (NeurIPS 2024 · 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.
