LLM Privacy Paradox: Balancing Data Utility...
summary
Experiment mixes LotR text with 5,000 synthetic PII generated via GPT then expanded with Gretel, and fine‑tunes models like GPT-2. When each record is repeated (100× duplication) the model memorizes and reproduces exact sequences (credit‑card–style outputs appear when a hyphen delimiter matches training patterns); single occurrences produce hallucinated but structurally similar data. Larger parameter models leak more. Data prep (de‑duplication, normalize chars, remove delimiters, dimension reduction), rate limiting and output sanitization materially reduce leaks.
AI-written summary · about
chapters (6)
- 0:00Introduction and research background
- 10:52Experiment setup and data preparation
- 22:50Key findings from experiments
- 38:23Mitigation strategies and defenses
- 1:04:46AI for vulnerability detection and fixing
- 1:22:56Q&A and conclusions
related talks
28mAnonymization in the LLM Era: No Room for Compromise
10mWhen Privacy Guarantees Meet Pre-Trained LLMs: A Case Study in Synthetic Data
24mHarnessing the Power of LLMs in Advanced Analytics
40mBalancing Privacy & Utility: Efficient PII Detection and Replacement in Textual Data
01h 17mTutorial: Privacy, Part I
01h 04mCan you trust your (large language) model?
17mPrivate Tuning of LLMs in Practice: From VaultGemma to Custom Fine-Tuning
21mFrom Data to Precision: Fine-Tuning LLMs for Proprietary Query Languages