Region-Aware Spatial-Frequency Transformation for Transferable Adversarial Attacks in Remote Sensing Object Recognition

Adversarial attacks provide an effective tool for exposing the vulnerabilities of deep neural networks. In practice, models are often deployed under black-box settings, making the transferability of adversarial examples generated on surrogate models a critical requirement. Existing methods commonly enhance transferability through input transformations during gradient computation. However, in remote sensing object recognition, targets are tightly coupled with complex backgrounds, which amplifies cross-model discrepancies due to spatially imbalanced attention. Moreover, most transfer-based attacks optimize perturbations by maximizing the classification loss, often inducing source-biased updates that overfit specific decision boundaries. To address these issues, we propose a region-aware spatial-frequency transformation (RASFT) in remote sensing object recognition. Specifically, a differentiated transformation strategy is applied to salient and background regions by jointly exploiting complementary spatial-domain geometric and frequency-domain perturbations through a learnable adaptive fusion mechanism. Moreover, an entropy regularization term is incorporated into the attack objective to encourage prediction uncertainty and mitigate source-model overfitting. Extensive experiments on FGSCR-42 and MTARSI datasets demonstrate that RASFT consistently outperforms state-of-the-art methods in terms of attack success rate and transferability across diverse model architectures.