Deep learning is a hot topic in computer vision and natural language processing. In many applications, the performance of deep neural networks (DNN) is better than traditional methods, and has been successfully applied to tasks such as modulation classification and radio signal representation learning. In recent years, studies have found that deep neural networks are vulnerable to adversarial attacks and lack robustness to "adversarial examples". Regarding the robustness of the neural network communication signal recognition algorithm, the author regards the data that has been attacked by the PGD as model-based data, and inputs the data into the neural network, which makes the signal recognition and classification results wrong. Then with the help of model-based defense algorithms, which are robust training algorithm and adversarial training algorithm, the experimental results after training show that both methods have good defense effects.