ALAT: Adversarial Label-guided Adversarial Training

Adversarial training is a widely used defense method in deep neural networks that enhances the model’s ability to detect perturbations and improve network robustness. Previous studies have assessed adversarial attack performance from various angles, leading to enhancements in adversarial training to bolster network robustness. However, while some research has explored the effectiveness of misclassified data in improving adversarial training, they have ignored the importance of the adversarial predicted labels. We observe that adversarial sample prediction labels often correspond to high probability categories in natural predictions. This paper proposes a new adversarial training method called Adversarial Label-guided Adversarial Training (ALAT). This method incorporates an additional regularization term that integrates adversarial prediction labels into the training process, guiding predictions closer to true labels and away from adversarial labels. Extensive experiments confirm its effectiveness.