Collaborative Evolutionary Multitasking Multiobjective Adversarial Attacks in Hyperspectral Image Classification

Adversarial attacks on hyperspectral image (HSI) classification has garnered significant attention. However, most of the existing adversarial attack methods tend to apply global perturbations to the HSI, ignoring the fact that different classes in the HSI have distinct preferences for the spectral bands, and fails to generate an approximate optimal perturbation for each class. Therefore, this paper establishes an evolutionary multitasking multiobjective adversarial attack model to generate a universal adversarial perturbation for each class. Then, we propose a novel collaborative evolutionary multitasking optimization (CoEMTO) method to optimize all tasks collaboratively. In CoEMTO, the optimal solution is the combination of solutions for all tasks, and a dynamic reward resource allocation strategy is designed to approximate the optimal solution for all tasks. In addition, perturbations are applied to the abundance rather than directly manipulating the pixel spectra to ensure the physical consistency of the generated adversarial perturbations. Experimental results demonstrate that the proposed CoEMTO algorithm can efficiently converge toward the Pareto front, and significantly outperforms other adversarial attack methods in terms of attack effectiveness.