Explore the concept of privacy amplification in data analysis through this 50-minute lecture by Yu-Xiang Wang from UC Santa Barbara. Delve into topics such as differential privacy, stochastic gradient descent, and sampling techniques. Examine the generic amplification lemma and its applications, as well as the technical aspects of sampling without replacement. Gain insights into various divergences and subsampling methods used in privacy-preserving data analysis. Understand the importance of privacy in the context of modern data science and learn about cutting-edge techniques for enhancing privacy in analytical processes.
Privacy Amplification by Sampling and Renyi Differential Privacy
Overview
Syllabus
Introduction
Agenda
Differential Privacy
Other mechanisms
SGD
Use Cases
Generic amplification lemma
Generic amplification bomb
Sampling without replacement
Intuition
Technical result
Analysis
Comparison
Data Structure
Results
Proof
Divergences
Subsampling
Taught by
Simons Institute
Related Courses
-
Cynthia Dwork: The Mathematics of Privacy
-
Federated Learning with Formal User-Level Differential Privacy Guarantees
-
Private Convex Optimization via Exponential Mechanism - Differential Privacy for Machine Learning
-
Gaussian Differential Privacy
-
Improving the Privacy-Utility Tradeoff in Differentially Private Machine Learning with Public Data
-
Gradients Look Alike: Sensitivity is Often Overestimated in DP-SGD
