To address the problems of high computational cost and catastrophic overfitting in the existing adversarial training algorithms for image classification neural network models, a learnable perturbation superposition adversarial training algorithm based on the generative network combined with small batch samples is proposed. During training, both the generative model and the target model undergo supervised training, utilizing the final perturbation for adversarial training to enhance the robustness of the target network. In the paper, experimental validation is carried out by combining two multi-classification datasets, CIFAR-10 and CIFAR-100, as well as three network architectures, ResNet18, WideResNet34, and VGG11. The experiments demonstrate that the model robustness under this algorithm is better than the baseline algorithms, and the computational efficiency is three times higher than that of the multi-step gradient projection adversarial training algorithm.