Profile-Then-Simulate: Can LLMs Faithfully Generate Differentially Private Synthetic...
summary
Profile-then-Simulate takes 2,000 users and 24M transactions, distills 1,600 users into DP-protected personas (all epsilon spent there) and seeds PersonaLedger (a Llama 3.370B model with rule engine) to simulate transactions, compared to direct DP synthesis via AIM. Direct synthesis consistently yields better fraud-detection utility and distributional fidelity across epsilons 1, 5, 10. LLM-generated data shows systematic priors: retired rate shifts ~50%→80% and transactions cluster 9am–2pm; binary features hold up but complex, high-cardinality patterns break.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 0:53Privacy-preserving financial data
- 2:19Profile-then-simulate approach
- 4:02Direct synthesis baseline
- 4:43Evaluation setup
- 5:49Utility results
- 7:35Fidelity analysis
- 8:17LLM bias root causes
- 9:41LLM faithfulness bottleneck
- 10:23Practical recommendations
- 12:11Future directions
related talks
55mFederated, Synthetic, Personalized: Heterogeneity Here or There?
10mWhen Privacy Guarantees Meet Pre-Trained LLMs: A Case Study in Synthetic Data
45mPractical Experience with Making Synthetic Data Differentially Private
01h 27mLLM Privacy Paradox: Balancing Data Utility...
01h 14mStatistical Inference and Privacy, Part III
32mThe Inadequacy of Offline LLM Evaluations: A Need to Account for Personalization in Model Behavior
16mCompute Engine Testing with Synthetic Data Generation
28mInsights, Not Secrets: How Synthetic Data Empowers Privacy