Provable defenses against adversarial examples via the convex outer adversarial polytope

We propose a method to learn deep ReLU-based classifiers that are provably robust against norm-bounded adversarial perturbations (on the training data; for previously unseen examples, the approach will be guaranteed to detect all adversarial examples, though it may flag some non-adversarial examples as well). The basic idea of the approach is to consider a convex outer approximation of the set of activations reachable through a norm-bounded perturbation, and we develop a robust optimization procedure that minimizes the worst case loss over this outer region (via a linear program). Crucially, we show that the dual problem to this linear program can be represented itself as a deep network similar to the backpropagation network, leading to very efficient optimization approaches that produce guaranteed bounds on the robust loss. The end result is that by executing a few more forward and backward passes through a slightly modified version of the original network (though possibly with much larger batch sizes), we can learn a classifier that is provably robust to any norm-bounded adversarial attack. We illustrate the approach on a toy 2D robust classification task, and on a simple convolutional architecture applied to MNIST, where we produce a classifier that provably has less than 8.4% test error for any adversarial attack with bounded $\ell_\infty$ norm less than $\epsilon = 0.1$. This represents the largest verified network that we are aware of, and we discuss future challenges in scaling the approach to much larger domains.