Increasing Confidence in Adversarial Robustness Evaluations

TL;DR

This paper introduces a test to identify weak adversarial attacks, revealing that many previous robustness claims are invalid, thereby aiming to improve the reliability of adversarial robustness evaluations.

Contribution

The authors propose a novel attack unit test that modifies neural networks to verify the strength of adversarial attacks, exposing weaknesses in prior evaluations.

Findings

Most previous defenses failed the new attack test

Stronger attacks successfully broke defenses that previously appeared robust

The test can reliably identify weak attack methods

Abstract

Hundreds of defenses have been proposed to make deep neural networks robust against minimal (adversarial) input perturbations. However, only a handful of these defenses held up their claims because correctly evaluating robustness is extremely challenging: Weak attacks often fail to find adversarial examples even if they unknowingly exist, thereby making a vulnerable network look robust. In this paper, we propose a test to identify weak attacks, and thus weak defense evaluations. Our test slightly modifies a neural network to guarantee the existence of an adversarial example for every sample. Consequentially, any correct attack must succeed in breaking this modified network. For eleven out of thirteen previously-published defenses, the original evaluation of the defense fails our test, while stronger attacks that break these defenses pass it. We hope that attack unit tests - such as ours - will be a major component in future robustness evaluations and increase confidence in an empirical field that is currently riddled with skepticism.