RSIAW: A Robust and Inconspicuous Watermarking Framework for Remote Sensing Images Using Adversarial Perturbations and Style Transfer

Remote sensing imagery is crucial for applications such as urban planning, resource management, and defense, yet it is increasingly vulnerable to data theft and tampering. Traditional protection methods, such as encryption and access control, are limited, especially with the rise of deep neural networks that can effectively remove visible watermarks. This paper presents RSIAW, a watermarking technique that combines adversarial perturbations and style transfer to enhance robustness and authenticity. Adversarial perturbations mislead watermark removal models, while style transfer improves visual coherence by integrating the watermark with the image texture. Evaluated on the NWPU-RESISC45 dataset, RSIAW demonstrates strong resistance to removal attacks while preserving image usability. The method's performance is assessed using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Root Mean Square Error (RMSE), yielding a PSNR of 0.984, SSIM of 42.59, and RMSE of 2.24, indicating high similarity and quality between the watermarked and original images. RSIAW outperforms baseline methods in resistance to removal attacks, achieving the highest SSIM (0.996) and PSNR (49.04) values with a watermark size of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$40\times 40$</tex> and opacity of 0.4.