计算机科学
数字水印
计算机视觉
人工智能
稳健性(进化)
图像(数学)
噪音(视频)
图像处理
模式识别(心理学)
图像分割
特征提取
扩散
复制保护
变换编码
离散余弦变换
可视化
各项异性扩散
算法
方案(数学)
作者
Shuren Qi,Wenying Wen,Mengxin Shen,Yushu Zhang,FengLei Fan
标识
DOI:10.1109/tmm.2026.3734777
摘要
Diffusion models have enabled highly realistic image generation, but their misuse also raises growing concerns about misleading visual content. Digital watermarking has emerged as a promising proactive solution for verifying AI-generated images in cooperative deployment settings. An effective watermarking method for diffusion models should ideally satisfy two key properties: imperceptibility and robustness. However, existing methods still face a clear trade-off between them: pixel-level approaches usually preserve image fidelity well but are vulnerable to regeneration-based removal, whereas content-level approaches are often more robust but may compromise reproducibility, pixel-level consistency, or deployment efficiency due to retraining. In this paper, we revisit diffusion watermarking from the perspective of embedding space design. We argue that a major difficulty lies in finding an embedding space that simultaneously captures components that are both visually insensitive and stable under common distortions. To address this issue, we propose a plug-and-play watermarking framework that embeds watermark information into the circular harmonic expansion of the initial noise in diffusion models. This embedding space provides two favorable properties for watermarking: amplitude-based invariance, which supports reliable detection under geometric transformations, and reconstructability with frequency selectivity, which helps preserve the original generation behavior while limiting perceptual distortion. Extensive experiments demonstrate that the proposed method achieves a strong robustness-imperceptibility trade-off without retraining the diffusion model. In particular, it shows strong robustness against natural distortions, geometric transformations, and regeneration-based expert attacks, while substantially improving imperceptibility over representative content-level baselines. Moreover, the method remains lightweight and broadly compatible with existing diffusion-model pipelines, making it practical for active AI-generated image verification.
科研通智能强力驱动
Strongly Powered by AbleSci AI