On Adversarial Robust Generalization of DNNs for Remote Sensing Image Classification

Deep neural networks (DNNs)-based deep learning is an important technical support in the task of remote sensing image classification. But DNNs are susceptible to adversarial attacks. Adversarial training is one of the effective ways to obtain robust models that can against such attacks. However, many adversarial training strategies can only defend against known attacks and perform poorly when faced with unknown attacks, namely, although the adversarial robustness of the model is enhanced, the generalization is reduced. To alleviate this phenomenon, we propose a new method to improve the adversarial robust generalization of DNNs for remote sensing image classification within the framework of adversarial training. Specifically speaking, we impose a Euclidean regularization constraint on the gradient of the adversarial loss function in order to achieve a narrowing of the robust generalization gap by enhancing the gradient alignment between clean and adversarial examples. In addition, we incorporate a label smoothing strategy into the adversarial training process, aiming to further improve the adversarial robustness of the model by reducing its sensitivity to subtle fluctuations. The combination of the above two strategies can not only improve the adversarial robustness of the model but also improve its adversarial robust generalization. Finally, in the case of two classical attack approaches FGSM and PGD, we validate the effectiveness and feasibility of our method through extensive experiments on two commonly used remote sensing image classification datasets NWPU-RESSC45, RSSCN7, and WHU-RS19, demonstrating its superiority over previous methods, particularly in the context of improving the adversarial robust generalization of DNNs.