Evaluating Robustness of Neural Networks with Mixed Integer Programming

TL;DR

This paper introduces a fast mixed integer programming-based method for verifying neural network robustness, enabling the analysis of larger networks and providing precise adversarial accuracy metrics.

Contribution

The authors develop a novel, computationally efficient verifier for piecewise-linear neural networks that significantly outperforms previous methods in speed and scale.

Findings

Verifier is 100-1000 times faster than previous methods.

Successfully verified robustness of larger convolutional networks.

Determined the exact adversarial accuracy of an MNIST classifier.

Abstract

Neural networks have demonstrated considerable success on a wide variety of real-world problems. However, networks trained only to optimize for training accuracy can often be fooled by adversarial examples - slightly perturbed inputs that are misclassified with high confidence. Verification of networks enables us to gauge their vulnerability to such adversarial examples. We formulate verification of piecewise-linear neural networks as a mixed integer program. On a representative task of finding minimum adversarial distortions, our verifier is two to three orders of magnitude quicker than the state-of-the-art. We achieve this computational speedup via tight formulations for non-linearities, as well as a novel presolve algorithm that makes full use of all information available. The computational speedup allows us to verify properties on convolutional networks with an order of magnitude more ReLUs than networks previously verified by any complete verifier. In particular, we determine for the first time the exact adversarial accuracy of an MNIST classifier to perturbations with bounded $l_\infty$ norm $\epsilon=0.1$: for this classifier, we find an adversarial example for 4.38% of samples, and a certificate of robustness (to perturbations with bounded norm) for the remainder. Across all robust training procedures and network architectures considered, we are able to certify more samples than the state-of-the-art and find more adversarial examples than a strong first-order attack.