Adversarial Examples Are Not Bugs, They Are Features

TL;DR

This paper argues that adversarial examples are due to non-robust features inherent in data, which are highly predictive but brittle, challenging traditional notions of robustness in machine learning.

Contribution

The paper introduces a theoretical framework for non-robust features, demonstrating their prevalence in datasets and linking adversarial vulnerability to data geometry and robustness definitions.

Findings

Non-robust features are highly predictive yet brittle.

Adversarial examples stem from non-robust features present in data.

Misalignment between human robustness notions and data geometry explains adversarial vulnerability.

Abstract

Adversarial examples have attracted significant attention in machine learning, but the reasons for their existence and pervasiveness remain unclear. We demonstrate that adversarial examples can be directly attributed to the presence of non-robust features: features derived from patterns in the data distribution that are highly predictive, yet brittle and incomprehensible to humans. After capturing these features within a theoretical framework, we establish their widespread existence in standard datasets. Finally, we present a simple setting where we can rigorously tie the phenomena we observe in practice to a misalignment between the (human-specified) notion of robustness and the inherent geometry of the data.