Statistical Inference and Privacy, Part III
summary
Three concrete recipes for differentially private synthetic data: smallDB enumerates all databases of size ~O(log|F|/α^2) and samples via the exponential mechanism (ε‑DP, accuracy α = Θ((log|F|/ε n)^{1/3}) but requires exponential time); multiplicative weights maintains a public histogram, answers “easy” queries from the histogram and updates on “hard” ones (ε‑DP, similar α bounds, runtime linear in the data universe and number of hard queries); private GANs clip gradients and add noise to discriminator updates (practically good, slow to train, no formal utility guarantees). Ongoing benchmarks include the NIST DP Synthetic Challenge on ACS data.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:23PrivacyPreserving Barriers
- 8:25Small DP
- 13:45Plunkett of weights
- 21:58No more hard queries
- 23:48Verification server
- 29:23Bayesian framework
- 35:13Privacy tools
- 38:33Competitions
- 41:23Additional Applications
- 44:20Synthetic Data
- 47:48Virtual RDC
- 50:38Multiple Implicits
- 55:03General Utility Measures
related talks
45mPractical Experience with Making Synthetic Data Differentially Private
48mSynthetic Data as an Enabler for Learning from Decentralized Private Data
01h 00mDifferential Privacy And The Complexity Of Simple Queries
36mGenerating realistic, differentially private data sets using GANs
19mPractical Considerations for Differential Privacy
01h 17mTutorial: Privacy, Part I
13mDPSynth: From Research to Production—Engineering Differentially Private Synthetic...
01h 00mDifferential Privacy: Fundamentals to Forefront I