Learning with privacy in consensus + obfuscation

Abstract

We examine the interplay between learning and privacy over multiagent consensus networks. The learning objective of each individual agent consists of computing some global network statistic, and is accomplished by means of a consensus protocol. The privacy objective consists of preventing inference of the individual agents' data from the information exchanged during the consensus stages, and is accomplished by adding some artificial noise to the observations (obfuscation). An analytical characterization of the learning and privacy performance is provided, with reference to a consensus perturbing and to a consensus-preserving obfuscation strategy.