Visible and infrared image fusion provides crucial support for high-level visual tasks by integrating textural details with thermal target information. However, the security of existing fusion models remains understudied. This paper proposes a novel black-box attack method based on a transfer strategy, addressing a critical gap in current research. By generating adversarial samples using multiple surrogate models, perturbations are injected into the input images, causing reduced brightness, blurred edges, and loss of detail, which result in significant visual quality degradation. Experiments conducted on four advanced fusion models across three datasets demonstrate that post-attack fusion quality metrics decline by an average of 14.27%. Further validation using object detection models reveals an average accuracy reduction of 11.94%, with substantial missed detections observed in some samples. This study establishes a critical evaluation benchmark for fusion model security and provides direction for the development of more robust fusion models.
Pengcheng Gao (Sun,) studied this question.