BACKGROUND ATTACK BASED ON VISUAL CONCEALMENT FOR REMOTE SENSING OBJECT DETECTION

Current research on adversarial attacks against remote sensing object detectors has made significant progress. However, most existing methods face a critical limitation in physical-world applications, as they are easily perceptible to the human eye. To address this issue, a Background Attack based on Visual Concealment (BAVC) framework is proposed to balance visual concealment and attack effectiveness. Specifically, a Natural Dual-Semantic Loss Function (NDSLF), which comprises local background style constraint and global background similarity constraint, is introduced to ensure that the generated adversarial background achieves both attack capabilities and visual camouflage. Additionally, a background-adaptive training strategy is developed, utilizing flexible adaptive transformations to generate adversarial backgrounds adapted to complex environments, thereby improving robustness across diverse scenarios. Extensive experiments and comparative analyses demonstrate that the proposed method achieves superior performance in both digital and physical-world settings.