Most deep neural networks are trained by gradient descent with first-order gradients. However, these networks can be highly sensitive to minor perturbations in input data, which may adversely affect ...