Reproducibility Study of "Attack-Resilient Image Watermarking Using Stable Diffusion"

TMLR Paper4324 Authors

Abstract:

This paper presents a reproducibility study and robustness evaluation of the paper ‘Attack Resilient Image Watermarking Using Stable Diffusion’ by Zhang et al. (2024), which proposes ZoDiac, a Stable Diffusion-based framework for attack-resilient image watermarking. While successfully replicating the original method’s core claims—achieving >90% watermark detection rate (WDR) against diffusion-based regeneration attacks and across MS-COCO, DiffusionDB, and WikiArt datasets—we identify critical vulnerabilities under adversarial
and geometrically asymmetric attack paradigms. Our extended analysis demonstrates that gradient-based adversarial perturbations reduce ZoDiac’s WDR, a threat model absent in prior evaluations. We also investigate rotationally asymmetric attacks achieving WDR below 65%. Additionally, we explore a new loss function to mitigate these limitations. Despite these enhancements, composite attacks combining adversarial noise with other methods reduce WDR to near-zero, exposing vulnerabilities through multi-stage offensive pipelines. Our implementation can be found on Anonymous Github.

Submission Length: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Hsuan-Tien_Lin1

Submission Number: 4324