Are Adversarial Examples a Tool to Improve Machine Learning Performance?

I still think that adversarial examples are important, but maybe not quite as much as I used to. We've started to find that there's a tradeoff between accuracy on adversarial examples and accuracy on clean examples. When we train against weak adversarial examples, MNIST classifiers get more accurate. It seems like when you confront a really strong adversary, you tend to have to give something up.