基于扩散模型的无人机遥感目标检测物理对抗攻击方法研究OA

Although deep neural networks have achieved significant advancements across a range of visual tasks,they continue to be vulnerable to adversarial attacks.Compared to digital-domain attacks,physical-world adversarial attacks pose greater threats.In the context of adversarial attacks on UAV remote-sensing image object detection,it's essential to maintain stable effectiveness under complex conditions,such as varying viewpoints,distances,and lighting conditions.Optimising attack methods must be fully considered in light of the dynamics and diversity of real-world imaging environments.Although existing physical-domain adversarial attack methods can degrade the performance of object detection models,they often rely solely on pixel-level local texture optimisation,resulting in monotonous adversarial texture patterns and limited adaptability.To address the aforementioned issues,this paper proposed a diffusion model-based physical adversarial attack method.The proposed approach employed a pre-trained diffusion model as the generator,leveraging both image and text priors to guide the generation of adversarial textures.Within a comprehensive physical attack framework,it enabled vehicle camouflage in UAV remote-sensing object-detection tasks.Experimental results demonstrate that the proposed method achieves high attack success rates and strong cross-model transferability across multiple object detection models,outperforming comparative methods in attack effectiveness and texture pattern diversity.