Overview
Explore the necessity of interaction in distributed private learning in this 21-minute IEEE conference talk presented at the 2017 IEEE Symposium on Security & Privacy. Delve into the local model for privacy, where data is randomized on individual devices before being sent to a server for aggregate statistics computation. Examine the challenges in convex optimization problems, including logistic regression, support vector machines, and Euclidean median, where current locally differentially-private algorithms require extensive interaction. Investigate new algorithms with reduced or no interaction, and analyze lower bounds on the accuracy of noninteractive algorithms. Gain insights into the potential separation between interactive and noninteractive approaches in distributed private learning, and understand the implications for data control and privacy in large-scale deployments.
Syllabus
Is Interaction Necessary for Distributed Private Learning?
Taught by
IEEE Symposium on Security and Privacy