Harnessing adversarial examples with a surprisingly simple defense

Ali Borji

I introduce a very simple method to defend against adversarial examples. The basic idea is to raise the slope of the ReLU function at the test time. Experiments over MNIST and CIFA…