Title: Differentially Private Distributed Convex Optimization via
Functional Perturbation

Speaker: Erfan Nozari (UC San Diego)

Abstract: A common and implicit assumption in the development of most
distributed coordination algorithms is the honesty and trustworthiness
of all the agents, upon which they can freely share their information
to achieve their common objective. Therefore, although security issues
of such algorithms have received considerable attention thus far, until
very recently, little work had been done to incorporate privacy
considerations into distributed coordination algorithms. In this talk,
I first review the notion of differential privacy and show how ``data''
can be deliberately contaminated with carefully chosen noise to
preserve its privacy when released. We will see that this notion is so
general that can be utilized in almost every context. Then, we will
focus on the particular problem of distributed convex optimization as
one of the most important multi-agent coordination problems. We discuss
and compare the two wide-spread methods of privacy preservation in this
scenario, namely, the perturbation of inter-agent messages or objective
functions, and this motivates our novel approach to this problem called
functional perturbation. We will then see how this approach can be used
to solve the private distributed optimization problem, and address the
issues and difficulties that may arise as a result.