Abstract
The rapid advancement of deep learning has significantly enhanced the performance of spectral image reconstruction, propelling its widespread application in real-world scenarios. However, existing research primarily focuses on optimizing reconstruction quality under ideal imaging conditions, lacking a systematic investigation into the stability of these models against adversarial perturbations. This gap has emerged as a critical bottleneck restricting their reliable deployment in practice. To address this, we delve into the network architectures and optimization mechanisms of spectral reconstruction models, and propose the first universal adversarial attack framework tailored for this task. Based on this framework, we conduct a comprehensive robustness evaluation. Extensive experiments reveal three key findings: First, spectral compressive imaging (SCI) models exhibit profound vulnerability to adversarial perturbations, where attacks can cause severe degradation or even irreversible damage to reconstruction quality. Second, the adversarial effects demonstrate a cross-band propagation characteristic, wherein perturbations applied to a single spectral band disrupt the reconstruction of other bands. Finally, compared to models based on the coded aperture snapshot spectral imaging (CASSI) system, those utilizing optical filter-based systems exhibit superior adversarial robustness. These findings provide critical insights and a fundamental reference for designing highly robust SCI architectures in the future.