Adversarial Attacks on Medical Hyperspectral Imaging Exploiting Spectral-Spatial Dependencies and Multiscale Features

ICLR 2026 Conference Submission24993 Authors

Keywords: medical hyperspectral, adversarial attack, spectral-spatial dependencie, multiscale features

Abstract: Medical hyperspectral imaging (HSI) represents a transformative innovation in diagnosing diseases and planning treatments by capturing detailed spectral and spatial features of tissues. However, the integration of deep learning into medical HSI classification has unveiled critical vulnerabilities to adversarial attacks. These attacks compromise the reliability of clinical applications, potentially leading to diagnostic inaccuracies and jeopardizing patient outcomes. This study identifies two fundamental reasons for the susceptibility of medical HSI models to adversarial manipulation: their reliance on local pixel dependencies, which are essential for preserving tissue structures, and their dependence on multiscale spectral-spatial features, which encode hierarchical tissue information. To address these vulnerabilities, we propose a novel adversarial attack framework specifically tailored to medical HSI. Our approach introduces the Local Pixel Dependency Attack, which exploits spatial relationships between neighboring pixels, and the MultiScale Information Attack, which perturbs spectral and spatial features across hierarchical scales. Experiments on the Brain and MDC datasets reveal that our method significantly reduces classification accuracy, particularly for critical tumor regions, while maintaining imperceptible perturbations. Compared to existing methods, the proposed framework highlights the unique fragility of medical HSI models and underscores the urgent need for robust defenses. This work highlights critical vulnerabilities in medical HSI models and demonstrates how leveraging local pixel dependencies and multiscale spectral-spatial features can guide the development of targeted defenses to enhance model robustness and clinical reliability.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 24993