SRAttack: Super-Resolution as an Attack on Invisible Image Watermarking

ICLR 2026 Conference Submission15637 Authors

Keywords: Wartermark, Attack, Super-Resolution

Abstract: Invisible image watermarking is an important technology for copyright protection, enabling the sharing and spread of invisible images. The study of watermark attacks motivates the development of more robust invisible image watermarking. An ideal attack should effectively erase the watermark, preserve the visual quality of the image, and be simple and accessible. Although existing methods can partially destroy watermarks, they often fail to preserve satisfactory visual quality of image. In this paper, we propose the Super-Resolution Attack (SRAttack), a simple attack framework that leverages super-resolution (SR) to overcome these limitations. The proposed attack method first pre-degrade the watermarked image, then use a SR model to reconstruct a high-quality, watermark-free version, and finally down-sample it back to original dimensions of cover image. SRAttack operates in a black-box, training-free manner, and can be readily instantiated. Extensive experiments demonstrate that SRAttack not only effectively destroys watermarks embedded by various advanced watermark algorithms, but also improves visual quality in contrast to existing attack methods. Our research findings reveal a new perspective on invisible image watermarking research, indicating the need to consider attacks aimed at improving visual quality rather than just reducing it. Our code will be available soon.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 15637