Robustness Optimization of Image Classification Model Based on Adversarial Training

As an important part of artificial intelligence, deep neural networks have been widely used in various fields, and show superior performance in image classification tasks. However, with the improvement of the performance of deep learning image classification models, these high-precision models are vulnerable to security threats against samples, resulting in misjudgments in the models, which threatens the robustness of artificial intelligence systems. In the image classification task, the adversarial robustness of the model is promoted mainly through the iteration of adversarial training methods, but the existing adversarial samples are quite different from human perception, and the defense has certain limitations. Therefore, this paper studies the robustness optimization of image classification model based on adversarial training, and further improves the robustness of the classification model from the perspective of adversarial attack and defense. Starting with neural networks and image classification, AT -AMP, an improved adversarial training algorithm based on the weight parameters of the smooth model, improves the robust generalization gap and robust overfitting problems of the native adversarial training AT by searching for the smoothing minimum of the model weight loss, and achieves better adversarial robustness based on the comparison of different data sets and different models. The accuracy of image classification is further improved.