Multi-Task Adaptive Knowledge Transfer for Black-Box Adversarial Attacks on Hyperspectral Images

Deep neural networks have substantially improved the performance of hyperspectral image classification, yet they remain vulnerable to adversarial attacks. Existing attack methods usually manipulate pixel spectra directly, ignoring the physical mixing mechanism of remote sensing imaging and potentially generating adversarial samples with limited physical consistency and interpretability. Moreover, balancing attack effectiveness and perturbation imperceptibility remains a challenging multi-objective optimization problem. To address these issues, this paper proposes an evolutionary multi-task multi-objective adversarial attack framework based on inter-task knowledge transfer. Instead of perturbing raw pixel spectra, the proposed method introduces perturbations into abundance maps obtained through spectral unmixing, thereby improving the physical plausibility of the generated adversarial samples. The generation of class-specific universal perturbations is formulated as a collaborative multi-task optimization problem. To solve this problem, we develop a Self-Adaptive Multi-Objective Multi-Factorial Evolutionary Algorithm for Adversarial Attacks (SAMO-MFEA-AA). By modeling the attack generation processes for different land-cover classes as distinct yet correlated optimization tasks, SAMO-MFEA-AA dynamically captures synergistic relationships among tasks. An asymmetric adaptive cooperation matrix is employed to regulate the intensity of knowledge transfer, allowing beneficial perturbation patterns to be shared across related classes while reducing the risk of negative transfer. Extensive experiments on the Indian Pines and Salinas datasets demonstrate that the proposed framework achieves competitive hypervolume performance and favorable solution diversity compared with existing multi-objective optimization algorithms. In adversarial attack scenarios, the proposed method achieves effective attack success rates against representative classification networks while maintaining the physical plausibility of abundance-space perturbations.