Describe how we could build the deep models that are relatively robust to adversarial attack? Briefly explain why the model trained in such way could boost the robustness.