Positional Attention-Augmented Adversarial Defense for Remote Sensing Image Super-Resolution
Nan Chen, Biao Zhang, Hongjie He · 2025
Single-image super-resolution plays a critical role in enhancing the resolution of low-quality remote sensing images for various computer vision applications. However, deep learning-based SISR methods are highly vulnerable to adversarial attacks, which can severely degrade super-resolved outputs while leaving the low-resolution inputs seemingly unaffected. While recent studies have explored the robustness of SISR models against such attacks, the integration of vision transformers into super-resolution frameworks introduces new challenges due to their unique patch-based and self-attention mechanisms. In this paper, we propose a Positional Attention-Augmented Adversarial Defense (PAAD) framework to enhance the robustness of ViT-based super-resolution models. we develop a defense mechanism that leverages positional attention to mitigate adversarial perturbations while preserving high-frequency details in super-resolved images. Our experiments demonstrate that the proposed PAAD framework significantly improves robustness of remote sensing image super-resolution against adversarial attacks.