The network obtained through adversarial training exhibits a significant degree of generalization error. To this end, we propose a novel adversarial training method called feature orthogonalization and improved mixup from the perspective of feature constraints. This adversarial training method forces different categories of features to tend towards orthogonality during the training process, enabling better separation of features from different categories. By imposing improved mixup constraints on image features, it can better aggregate features from the same category. These two techniques do not require the introduction of additional parameters and hyperparameters. By integrating these two techniques into the adversarial training process, state-of-the-art (SOTA) effects can be achieved. Extensive experiments on CIFAR10 and SVHN datasets using Resnet18 and Vgg16 networks have demonstrated the effectiveness of our approach.