Differential Privacy with Elliptical Stochastic Processes
summary
Proposal: add noise drawn from an elliptical stochastic process to a function estimate, calibrate scale sigma by the RKHS norm of the change in the estimate, and get finite‑dimensional ε‑DP if the process density tail function satisfies a boundedness condition. Constructions using multivariate t processes work, but a multivariate Laplace noise fails for any dimension >1. A sharper negative claim is that no elliptical noise yields true ε‑DP for a truly infinite‑dimensional summary, though (ε,δ)‑DP remains achievable.
AI-written summary · about
chapters (6)
- 0:00Start
- 7:00Functional Data Example
- 13:26Multivariate Elliptical Distributions
- 15:55Elliptical Stochastic Processes
- 17:54Sensitivity
- 24:10Examples
related talks
01h 14mTutorial: Privacy, Part II
31mWhen Existing Techniques in Linear Regression Preserve Differential Privacy
35mDifferential Privacy Made Practical
21mDimension-free Private Mean Estimation for Anisotropic Distributions
09mFrom Robustness to Privacy and Back
30mAccuracy First: Selecting a Differential Privacy Level for Accuracy-Constrained...
01h 07mStatistical Inference and Privacy, Part I
01h 02mLarge-Scale Private Learning, Part I