Current browse context:
cs.LG
Change to browse by:
References & Citations
Computer Science > Machine Learning
Title: Measuring Neural Net Robustness with Constraints
(Submitted on 24 May 2016 (v1), last revised 16 Jun 2017 (this version, v2))
Abstract: Despite having high accuracy, neural nets have been shown to be susceptible to adversarial examples, where a small perturbation to an input can cause it to become mislabeled. We propose metrics for measuring the robustness of a neural net and devise a novel algorithm for approximating these metrics based on an encoding of robustness as a linear program. We show how our metrics can be used to evaluate the robustness of deep neural nets with experiments on the MNIST and CIFAR-10 datasets. Our algorithm generates more informative estimates of robustness metrics compared to estimates based on existing algorithms. Furthermore, we show how existing approaches to improving robustness "overfit" to adversarial examples generated using a specific algorithm. Finally, we show that our techniques can be used to additionally improve neural net robustness both according to the metrics that we propose, but also according to previously proposed metrics.
Submission history
From: Osbert Bastani [view email][v1] Tue, 24 May 2016 02:18:21 GMT (719kb,D)
[v2] Fri, 16 Jun 2017 11:58:51 GMT (817kb,D)
Link back to: arXiv, form interface, contact.