Increasing Confidence in Adversarial Robustness Evaluations
This addresses the challenge of skepticism in empirical robustness evaluations for the machine learning community, offering a tool to improve confidence in defense claims.
The paper tackles the problem of unreliable adversarial robustness evaluations in deep neural networks by proposing a test to identify weak attacks, revealing that 11 out of 13 previously-published defenses had flawed evaluations.
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.