Empirical Privacy Metrics: The Bad, the Ugly… and the Good, Maybe?
summary
Synthetic-data vendors often report similarity-based metrics that compute a distance-to-train vs distance-to-test and flag overfitting, but those numbers are easy to game, vary wildly with random seeds, assume an unrealistic attacker who tries to link synthetic points to real ones, and only report average risk that hides outliers. Better practice measures realistic attacker behavior, reports per-record or worst-case leakage, treats empirical metrics as alerts not guarantees, and pairs them with provable approaches like differential privacy to bound worst-case risk.
AI-written summary · about
chapters (6)
- 0:00The synthetic data myth
- 1:51Similarity-based metrics
- 3:53Flaws in privacy metrics
- 7:06The dishonesty of framing
- 8:46Industry incentives
- 10:50Path toward improvement
related talks
01h 19mThe Synthetic Data Spectrum: Where Does Anonymisation Start and Privacy-preserving end?
28mInsights, Not Secrets: How Synthetic Data Empowers Privacy
10mWhen Privacy Guarantees Meet Pre-Trained LLMs: A Case Study in Synthetic Data
47mDifferential Privacy from a Statistical Perspective–Obtaining Valid Inferences...
20mSynthetic Data for Privacy, AI Fairness and Democratizing Data Access
25mThe Importance of Synthetic Data in Data-Centric AI
45mSynthetic Data for Machine Learning Applications
46mThe Measure and Mismeasure of Fairness