The Limitations of Model Uncertainty in Adversarial Settings
arXiv:1812.02606
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.
arXiv commentAccepted to the Bayesian Deep Learning Workshop 2019 at NeurIPS. For longer version with more background, refer to previous version
References in corpus (6)
- Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms
- Wild Patterns: Ten Years After the Rise of Adversarial Machine Learning
- Transferability in Machine Learning: from Phenomena to Black-Box Attacks using Adversarial Samples
- Understanding Measures of Uncertainty for Adversarial Example Detection
- Adv-BNN: Improved Adversarial Defense through Robust Bayesian Neural Network
- Adversarial Robustness Toolbox v1.0.0
Cited by in corpus (7)
- Adversarial Attack and Defense on Point Sets
- Adversarial Robustness Guarantees for Classification with Gaussian Processes
- Bayesian Inference with Certifiable Adversarial Robustness
- LiBRe: A Practical Bayesian Approach to Adversarial Detection
- Adversarial Robustness Guarantees for Gaussian Processes
- Efficient and Transferable Adversarial Examples from Bayesian Neural Networks
- The Effect of Prior Lipschitz Continuity on the Adversarial Robustness of Bayesian Neural Networks