On Convex Stochastic Variance Reduced Gradient for Adversarial Machine Learning

Abstract

We study the finite-sum problem in an adversarial setting using stochastic variance reduced gradient (SVRG) optimization in a distributed setting. Here, a fraction of the workers are assumed to be Byzantine that exhibit adversarial behavior by providing arbitrary data. We propose a robust scheme to combat the actions of Byzantine adversaries in this setting, and provide rates of convergence for the convex case. This is the first study of SVRG in an adversarial setting.