Adversarial system of one-pixel attack for hyperspectral image classification

Abstract

The introduction of deep learning arouses great success in hyperspectral image (HSI) classification. However, there is a growing concern about adversarial attacks on deep models, which can significantly impact the performance of HSI classification through imperceptible perturbations. In this paper, we first introduce a semi-black-box attack method to hyperspectral domain, one-pixel attack, which fools the HSI classifier by modifying only one pixel with high intensity. It is verified on standard datasets and a well-known HSI classification model, HybridSN. An adversarial system based on one-pixel attack (OPAS) is then proposed by integrating adversarial training into model training. The system not only improves the robustness of the classification model but also enhances the model’s ability to defend against other adversarial attacks. Experimental results demonstrate show that OPAS can achieve nearly 100% defense effectiveness against some basic adversarial attacks. Additionally, an HSI adversarial map is created to visualize one-pixel attack. Through it can be observed the characteristic that one-pixel attack can attack classification boundaries, and this property is further verified in hyperspectral field.