EVENT DOSSIER
To address concerns about false information caused by text-guided image editing, the study proposed the AngelFingerprint watermark framework. This method integrates LoRA into the diffusion model, directly embedding the CLIP text used for editing into the weights of the embedding model. The extractor recovers this embedding from pixels to interpret the modified content. Its core consists of two technologies: speed-aligned anchors to maintain editing quality, and a dedicated frequency filter to ensure that the watermark is invisible but recoverable and robust. On the MagicBrush dataset, the extractor achieved an Top-1 accuracy of 86% in 200 types of prompt retrieval, significantly outperforming the 20% accuracy of the prompt reversal method.
针对文本引导图像编辑引发的虚假信息担忧，AngelFingerprint 提出了一种可追溯、可解释且白盒隐蔽的水印框架。该方法将 LoRA 集成至扩散模型中，直接将编辑提示的 CLIP 文本嵌入植入模型权重，并通过提取器仅从像素恢复该嵌入以解释修改内容。其核心包含两个技术：速度对齐锚点用于保持编辑质量，专用频率滤波器确保水印不可见但可恢复且鲁棒。在 MagicBrush 数据集上，该提取器在 200 种提示词检索中达到 86% 的 Top-1 准确率，显著优于提示词反转方法的 20%。