IOPA-FracAT: Research on Improved One-pixel Adversarial Attack and Fractional Defense in Hyperspectral Image Classification

The introduction of deep learning arouses great success in hyperspectral image (HSI) classification. However, adversarial attacks on the deep model that substantially affect the HSI classification performance by some imperceptible perturbations have achieved significant concerns. In this paper, we propose an improved semi-black-box attack method called improved one-pixel attack (IOPA), which focuses on deceiving HSI classifiers by modifying only one pixel with high intensity in high-dimensional HSI data. Building on this insight, we incorporate the adversarial examples generated by IOPA into the training set and further introduce a fractional-order adversarial training (Frac-AT) defense method. Frac-AT, with the memory properties of fractional-order optimization, better enhances the robustness of neighboring boundary samples, outperforming traditional adversarial training. To visualize the attack effects of IOPA and the defense effects of Frac-AT, we create a hyperspectral adversarial map. The experimental results demonstrate that IOPA outperforms existing state-of-the-art attack methods for HSIs. Moreover, implementing Frac-AT based on IOPA enables HybridSN to enhance the defense capability of the model against various adversarial attacks while also improving the robustness of boundary samples, thereby enhancing overall accuracy.