ARMOR: Manifold-oriented training for adversarially robust aerial object detection under data scarcity

ARMOR defends aerial object detection against physically universal adversarial patch attacks under data scarcity through manifold-oriented training: it reuses detection annotations, masks the image background, and injects random patches onto objects to enhance feature robustness. In low-data experiments covering query-free transfer and defense-aware attacks, ARMOR maintains clean model confidence above 0.90, and improves adversarial robustness by up to 0.32 over state-of-the-art defenses; physical experiments with printed patches confirm that the gains hold in real deployment.