Mixing normal images and adversarial images when training CNNs