.png){kind=icon}
.png){kind=icon}
Return
Invisible watermarking framework for unlearned diffusion model in online service
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neunet.2025.108477.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In recent years, diffusion models have made significant progress in image generation in the online service. We can query for an online model and download the results. However, diffusion models may lead to copyright disputes and ethical concerns. Specifically, generated content might unintentionally use copyrighted data without proper authorization; and the model may be misused to create misleading, harmful, or even illegal content. To address these pressing challenges, two technologies may needed: watermarking serves as a crucial tool for copyright protection in such models; and machine unlearning can be applied for delete harmful or illegal contents. However, these two technologies may have conflict when applying in the model. The processing of unlearning in diffusion models can inadvertently compromise existing watermarking techniques, affecting the invisibility, robustness, and overall effectiveness of watermarking. To simultaneously meet the demands of both unlearning and watermarking, this paper proposes an advanced invisible watermarking technique within the framework of unlearned diffusion models. Our approach ensures that the removal of unlearned data does not degrade the quality of the generated content while preserving the stealthiness, detectability, and resilience of the watermark. We have designed an optimized noise generation process that allows the watermark to be effectively embedded without interfering with the unlearning operation, ensuring its detectability and reliability. In addition, we improve the watermark generator to dynamically adapt to changes in the feature distribution of the model, allowing for robust, efficient, and adaptive watermark embedding.
Journal
IF:
6.3
Papers:
7.8K
Citations:
3.0W