对抗攻击下的高光谱图像分类

Hyperspectral Image Classification under Adversarial Attack

摘要: 近些年来深度学习在高光谱图像(HSI)分类领域取得了重大的进展，但是在面对对抗干扰时，神经网络所表现的脆弱性不容忽视。生成的对抗样本与干净样本相比，人眼几乎无法察觉，但是大多先进的深度学习模型可能会受对抗样本的愚弄，从而分类预测时误判。为了针对这一问题，本文提出了一种分层特征引导上下文网络(HFGCNet)。本文发现，通过利用高阶特征指导低阶特征学习的方法学习可以增强多尺度特征融合的有效性，能更好地提取出HSI所包含的全局上下文信息，为上下文网络的损失分摊模块提供一个更好的输入，显著地提高了神经网络面对对抗干扰时的鲁棒性。在两个公开HSI数据集上进行对比实验，结果表明与其他先进的深度学习模型相比，本文所提出的方法更具抵抗力。

Abstract: Deep learning has made significant progress in the field of hyperspectral image (HSI) classification in recent years, but the vulnerability exhibited by neural networks in the face of adversarial interference cannot be ignored. The generated adversarial samples are almost undetectable to the human eye compared to clean samples, but most advanced deep learning models may be fooled by the adversarial samples and thus misclassify the classification prediction. To address this problem, we propose a hierarchical feature-guided context network (HFGCNet). We find that learning by utilizing higher-order features to guide the learning of lower-order features enhances the effectiveness of multi-scale feature fusion, better extracts the global contextual information contained in the HSI, provides a better input to the loss apportionment module of the context network, and significantly improves the robustness of the neural network in the face of adversarial interference. Comparative experiments on two publicly available HSI datasets show that the method proposed in this paper is more resistant compared to other state-of-the-art deep learning models.