Adversarial example classification algorithm based on generative self-supervised learning

Adversarial examples are often regarded as a threat to the robustness of deep learning models, and various defense techniques such as adversarial training have been developed to mitigate the impact of adversarial examples on label prediction. However, the various existing adversarial training reduces the generalization accuracy of the classification network, resulting in a reduction in its classification effect on the original examples. Therefore, an adversarial example classification algorithm based on generative self-supervised learning is proposed. Through self-supervised learning, the generative model can be trained to obtain the potential features of image data, and this model performs feature screening on adversarial examples. After that, the information useful for classification is fed back to train the classification model. Finally, joint learning is carried out to complete the end-to-end global training, and further improves the generalization accuracy of the classification model. Experimental results on MNIST, CIFAR10, and CIFAR100 datasets show that compared with standard training, the proposed algorithm increases the classification accuracy by 0.06%, 1.34%, and 0.89%, respectively, reaching 99.70%, 84.34%, and 63.65%. The result shows that it overcomes the inherent shortcomings of traditional adversarial training reducing the generalization performance of the model, and further improves the accuracy of the classification network.