Adversarial Training Based on Latent Distribution Calibration

Deep neural networks show vulnerability to adversarial examples, so it is necessary to study adversarial defense techniques in depth to improve the robustness of deep learning models. Most of the existing defense methods use specific adversarial attack algorithms and parameter settings that depend on experience, and their defense capabilities against unknown attacks are insufficient. In this paper, we propose a defense framework based on adversarial training and adversarial purification. Specifically, we first design an adversarial purifier based on latent distribution calibration. Then, with the help of the purifier, the policy network learns to generate more adversarial examples to maximize the training loss, while the target network learns to minimize the training loss on adversarial examples. We conduct comparative experiments on standard datasets such as CIFAR-10 and CIFAR-100 with advanced adversarial training methods. The results show that the proposed adversarial defense framework exhibits competitive robust accuracy and is more suitable for dealing with unknown adversarial attacks.