The Limitations of Model Uncertainty in Adversarial Settings

TL;DR

This paper explores the limitations of Bayesian neural network uncertainty measures in adversarial settings, revealing that confidence and uncertainty can be misleading even when models are wrong, highlighting challenges in model reliability.

Contribution

The study demonstrates that Bayesian uncertainty measures can be unreliable under adversarial perturbations and compares them with Gaussian process classifiers to analyze their behavior.

Findings

Uncertainty and confidence can be misleading in adversarial scenarios.

Bayesian neural network uncertainty measures can be unsuspicious even when predictions are wrong.

Differences exist between features influencing uncertainty and confidence in most tasks.

Abstract

Machine learning models are vulnerable to adversarial examples: minor perturbations to input samples intended to deliberately cause misclassification. While an obvious security threat, adversarial examples yield as well insights about the applied model itself. We investigate adversarial examples in the context of Bayesian neural network's (BNN's) uncertainty measures. As these measures are highly non-smooth, we use a smooth Gaussian process classifier (GPC) as substitute. We show that both confidence and uncertainty can be unsuspicious even if the output is wrong. Intriguingly, we find subtle differences in the features influencing uncertainty and confidence for most tasks.