Optimal Attacks against Multiple Classifiers

Juan C. Perdomo, Yaron Singer

Abstract: We study the problem of designing provably optimal adversarial noise algorithms that induce misclassification in settings where a learner aggregates decisions from multiple classifiers. Given the demonstrated vulnerability of state-of-the-art models to adversarial examples, recent efforts within the field of robust machine learning have focused on the use of ensemble classifiers as a way of boosting the robustness of individual models. In this paper, we design provably optimal attacks against a set of classifiers. We demonstrate how this problem can be framed as finding strategies at equilibrium in a two player, zero sum game between a learner and an adversary and consequently illustrate the need for randomization in adversarial attacks. The main technical challenge we consider is the design of best response oracles that can be implemented in a Multiplicative Weight Updates framework to find equilibrium strategies in the zero-sum game. We develop a series of scalable noise generation algorithms for deep neural networks, and show that it outperforms state-of-the-art attacks on various image classification tasks. Although there are generally no guarantees for deep learning, we show this is a well-principled approach in that it is provably optimal for linear classifiers. The main insight is a geometric characterization of the decision space that reduces the problem of designing best response oracles to minimizing a quadratic function over a set of convex polytopes.

Keywords: online learning, nonconvex optimization, robust optimization

TL;DR: Paper analyzes the problem of designing adversarial attacks against multiple classifiers, introducing algorithms that are optimal for linear classifiers and which provide state-of-the-art results for deep learning.