Multi-Scale Feature Dispersion: Towards Occlusion-Resilient Adversarial Patches for UAV Perception

Deep learning-based perception is fundamental to Unmanned Aerial Vehicles (UAVs), yet it remains vulnerable to physical adversarial patches. Existing methods for generating adversarial patches typically rely on localized key features, making them highly fragile under partial occlusion, a common scenario in UAV operations due to environmental obstruction and viewpoint variation. To address this limitation, we propose Multi-Scale Feature Dispersion (MSFD), an information-theoretic framework for generating robust adversarial patches under incomplete observations. MSFD maximizes information entropy to promote statistically uniform perturbations, while spatial autocorrelation introduces structural redundancy to preserve attack effectiveness when critical regions are occluded. Additionally, a multi-scale consistency constraint ensures robustness across varying flight altitudes. Experiments on the VisDrone dataset and in high-fidelity AirSim environments demonstrate that MSFD achieves an attack success rate (ASR) of 45.6% under 50% occlusion, whereas existing methods degrade to near-zero performance. These results highlight the importance of feature dispersion in adversarial robustness and provide a principled approach for evaluating perception security in real-world UAV scenarios.