Abstract
In recent years, there has been significant research focusing on addressingsecurity concerns in single-modal person re-identification (ReID) systems thatare based on RGB images. However, the safety of cross-modality scenarios, whichare more commonly encountered in practical applications involving imagescaptured by infrared cameras, has not received adequate attention. The mainchallenge in cross-modality ReID lies in effectively dealing with visualdifferences between different modalities. For instance, infrared images aretypically grayscale, unlike visible images that contain color information.Existing attack methods have primarily focused on the characteristics of thevisible image modality, overlooking the features of other modalities and thevariations in data distribution among different modalities. This oversight canpotentially undermine the effectiveness of these methods in image retrievalacross diverse modalities. This study represents the first exploration into thesecurity of cross-modality ReID models and proposes a universal perturbationattack specifically designed for cross-modality ReID. This attack optimizesperturbations by leveraging gradients from diverse modality data, therebydisrupting the discriminator and reinforcing the differences betweenmodalities. We conducted experiments on three widely used cross-modalitydatasets, namely RegDB, SYSU, and LLCM. The results not only demonstrate theeffectiveness of our method but also provide insights for future improvementsin the robustness of cross-modality ReID systems.