Date of Award

2020

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Graduate Group

Applied Mathematics

First Advisor

Aaron L. Roth

Abstract

Differential privacy has seen remarkable success as a rigorous and practical for-

malization of data privacy in the past decade. But it also has some well known

weaknesses, lacking comprehensible interpretation and an accessible and precise

toolkit. This is due to the inappropriate (ε, δ) parametrization and the frequent

approximation in the analysis. We overcome the difficulties by

1. relaxing the traditional (ε, δ) notion to the so-called f -differential privacy

from a decision theoretic viewpoint, hencing giving it strong interpretation,

and

2. with the relaxed notion, perform exact analysis without unnecessary approx-

imation.

Miraculously, with the relaxation and exact analysis, the theory is endowed with

various algebraic structures, and enjoys a central limit theorem. The central limit

theorem highlights the role of a specific family of DP notion called Gaussian Dif-

ferential Privacy. We demonstrate the use of the tools we develop by giving an

improved analysis of the privacy guarantees of noisy stochastic gradient descent.

Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS