Robustness-aware decoupling framework for adversarial detection in remote sensing images

Adversarial example detection is essential for ensuring the reliability of deep neural networks in remote sensing (RS) image analysis. However, existing methods often neglect the uneven impact of adversarial perturbations on feature representations, leading to limited generalization across attack settings. This paper proposes a robustness-aware decoupling (RAD) framework for adversarial detection in RS images. The framework disentangles features into perturbation-insensitive robust and perturbation-sensitive non-robust components and performs comparative pattern analysis for detection, thereby enabling detection decisions to be driven by intrinsic feature discrepancies rather than attack-specific perturbation patterns. Specifically, a hierarchical rotation-enhanced representation (HRER) module captures scale-diverse and orientation-aware features aligned with RS characteristics, while an adaptive robustness-aware feature decoupler (ARFD) conducts element-wise robustness evaluation to explicitly separate robust and non-robust features. Finally, a differential bidirectional Mamba head (DBMH) models global visual context to compare the decoupled representations, enhancing sensitivity to subtle adversarial-induced variations. Experiments on UCM and AID show that RAD achieves average detection accuracies of up to 92.71% and 90.52%, respectively, across different backbones, while maintaining strong cross-attack generalization, with average accuracies of 87.65% on UCM with ResNet-50 and 92.14% on AID with MobileNet-V3. These results demonstrate that RAD provides a robust and generalizable solution for adversarial detection in RS images.