Improving Adversarial Robustness by Reconstructing Interclass Relationships

Deep neural network models perform well on image classification tasks, but they are highly susceptible to adversarial samples that add tiny perturbations and output incorrect classification results. The existence of adversarial samples seriously threatens the security of the model itself, in order to mitigate the threat posed by adversarial samples, researchers and scholars have proposed a number of adversarial sample defense methods, which are currently considered to be one of the most effective adversarial defense methods is adversarial training. we redefine the loss in standard adversarial training as a new loss in order to enable the model to better learn the interclass relationships of the examples, which consists of three items: natural categorization loss, target-related loss, and target-irrelevant loss. Through experimental analysis, the target-irrelevant loss has a greater impact on the robustness of the model, and the combination of the three can better improve the robustness of the model, and the DRAT algorithm is further proposed. The experimental results show that the method can further improve the adversarial robustness of the model, and the robust accuracy under multiple attacks outperforms other classical adversarial training methods.