Improving Global Adversarial Robustness Generalization With Adversarially Trained GAN