The Limitations of Model Uncertainty in Adversarial Settings
This work addresses security vulnerabilities in machine learning models for adversarial robustness, but it is incremental as it builds on existing uncertainty methods.
The paper investigates how Bayesian neural networks' uncertainty measures behave under adversarial attacks, showing that both confidence and uncertainty can appear normal even when the output is incorrect, with subtle differences in feature influences observed.
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.