Speaker:
Sewoong Oh
Title:
Data processing inequality for differential privacy and applications
Abstract:
We provide a hypothesis testing interpretation to differential privacy and derive natural
forward and reverse data processing inequalities. These inequalities are very useful
in deriving tight impossibility results, as demonstrated by the following two applications:
composing multiple queries and multi-party computation. The impossibility results hold for
arbitrary parameter settings (privacy levels, number of queries, etc) and for both standard
and approximate differential privacy settings. Further, these impossibility results cannot
be improved upon.