PLKA-Net: Pyramid-Enhanced Large Kernel Attention Network for Robust Hyperspectral Image Classification Against Adversarial Attacks

Deep neural networks (DNNs) made significant research advances in the classification of hyperspectral images (HSI). However, adversarial perturbations can easily mislead DNNs by subtly altering input data, resulting in incorrect classification outcomes with high confidence. Current approaches suffer dual limitations: Firstly, constrained convolutional receptive fields compromise the modeling of global feature dependencies, reducing structural pattern recognition capability under perturbations. Secondly, multi-scale feature conflicts emerge: shallow layers overemphasize local textures while deep layers tend to discard crucial high-frequency information. These issues collectively degrade the generalization capacity of feature fusion. In light of these problems, this paper proposes a pyramid-enhanced large kernel attention network (PLKA-Net). This network designs a large kernel convolution (LKC) attention module to expand the receptive field range to capture global features, and fuses channel attention to achieve key region focusing. The pyramid-enhanced fusion (PEF) module adopts a pyramid pooling structure to aggregate multi-level spatial contextual information through multi-scale pooling and deeply fuses it with the high-resolution features of the LKC attention module. Extensive experiments on three benchmark HSI datasets indicate that PLKA-Net surpasses current state-of-the-art methods under various adversarial attacks, achieving stable classification accuracy alongside effective defense performance.