[Theory] Fwd: Theory Lunch 2023-01-25T18:30:00.000Z

[Antares Chen]

---------- Forwarded message ---------
From: TheoryBot (via Airtable) <noreply+automations at airtableemail.com>
Date: Wed, Jan 25, 2023 at 7:54 AM
Subject: Theory Lunch 2023-01-25T18:30:00.000Z
To: <antaresc at uchicago.edu>
CC: <ctkang at uchicago.edu>


*Jinshuo Dong (Northwestern): Exact Privacy via Isoperimetry*

https://uchicago.zoom.us/j/91616319229?pwd=dDdXQnFXeGNubFRkZy9hTDQrcWlXdz09

Description: Differential privacy has been a popular approach for formal
data privacy. An important part of a successful practice is the
characterization of the exact amount of privacy offered by algorithms.
Existing approaches for private convex optimization involve noisy gradient
descent with various artificial tweaks and tricks, which make it
challenging to compute the exact amount of privacy. The line of work by
Gopi et al (arXiv:2203.00263 and 2207.08347) suggested a beautiful and
intriguing path via the isoperimetric inequality for log-concave measures,
partly built on my own work. I will discuss how the two concepts are
related and do some of the most intriguing proofs, with the hope of getting
comments and ideas that can help us (mostly me) better understand the
underlying math.

Yours always,

TheoryBot
Sent via Automations on [image: Airtable]

©2023 Airtable
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/theory/attachments/20230125/cc333d3b/attachment-0001.html>