Towards adversarial robustness with 01 loss neural networks