Detecting Adversarial Examples for Hyperspectral Image Classification Via Mutation Testing

With the widespread application of remote sensing technology in various fields, convolutional neural networks (CNNs) have exhibited remarkable performance in the field of hyperspectral image (HSI) classification. However, CNNs also introduce security concerns when demonstrating their accurate classification capabilities. By adding imperceptible and carefully designed adversarial perturbations to input images, CNN models can produce highly confident incorrect results. These images, which have adversarial perturbations added to them, are known as adversarial samples. In this work, we propose a mutation testing-based method for detecting adversarial examples. In our experiments, we employ different attack algorithms to deceive multiple high-accuracy HSI recognition models trained on multiple HSI datasets, and our method aims to determine whether the generated adversarial examples are normal or adversarial. We evaluate our method on the Indian Pines, Salinas, and University of Pavia datasets. The results demonstrate that our method can effectively detect adversarial examples generated by state-of-the-art attack methods.