Joint Adversarial Defense Based on Gradient Regularization Distillation

In order to solve the adversarial defense issue for image classification, several researches have suggested different strategies to improve the robustness of neural networks. However, these techniques encounter challenges such as insufficient adaptability to sophisticated attacks and high computational overhead. In this work, we propose JGRD to employ gradient regularization and defense distillation strategies to improve the robustness against imperceptible perturbations. Furthermore, we jointly train the projection gradient descent strategy and class-specific maximization strategy to improve the accuracy on sophisticated adversarial samples and reduce the time cost of adversarial training. Experiments on the Cifar10 show that JGRD not only outperforms baselines, but also reduces the time cost of adversarial training.