When Privacy Guarantees Meet Pre-Trained LLMs: A Case Study in Synthetic Data
summary
Analysis finds a differentially private synthetic financial dataset on Hugging Face with about 1000 synthetic records derived from ~14,000 SEC filings, claiming epsilon = 8 but omitting delta and using document-level privacy units. Simple NER (spaCy) extracts ~100× more names than some transformer baselines, clustering yields eight unexpected clusters, and lexical signatures let Google recover original filings. Recommendations include membership tests, canary insertion, and a six-item preflight checklist.
AI-written summary · about
chapters (8)
- 0:00Synthetic data privacy risks
- 1:10Synthetic data promises and challenges
- 2:36Differentially private financial dataset
- 4:24Black-box threat model and privacy attacks
- 6:28Clustering and distribution attacks
- 7:49Membership inference and canary attacks
- 9:13Key takeaways for synthetic data
- 10:14Q&A
related talks
01h 19mThe Synthetic Data Spectrum: Where Does Anonymisation Start and Privacy-preserving end?
28mInsights, Not Secrets: How Synthetic Data Empowers Privacy
01h 27mLLM Privacy Paradox: Balancing Data Utility...
20mSynthetic Data for Privacy, AI Fairness and Democratizing Data Access
48mSynthetic Data as an Enabler for Learning from Decentralized Private Data
01h 14mStatistical Inference and Privacy, Part III
33mPractical Data Privacy
45mPractical Experience with Making Synthetic Data Differentially Private