Adversarial defense via perturbation-disentanglement in hyperspectral image classification

In recent years, deep neural networks (DNNs) have been widely used in hyperspectral image (HSI) classification. However, it has a strong vulnerability to crafted adversarial examples. Therefore, defense against adversarial examples is an urgent problem to be solved. To date, most defense methods are difficult to defend against unknown attacks. In this paper, we propose a perturbation-disentanglement-based adversarial defense method (PD-Defense) to protect HSI classification networks from unknown attacks. In the proposed method, the adversarial examples are decoupled into attack-invariant features and perturbation features, and the defense is conducted on the attack-invariant feature to defend against unknown attacks. Extensive experiments are performed on two benchmark HSI datasets, including PaviaU and HoustonU 2018. The results indicate that the proposed PD-Defense method achieves an excellent defense performance compared to four state-of-the-art defense methods.