首页 > AI前沿 > Distortion-Free Semantic Watermarks for Black-Box Generative Models

Distortion-Free Semantic Watermarks for Black-Box Generative Models

arXiv机器学习 2026-11-18 02:49 6 阅读 查看原文
用户现在需要把给定的新闻正文排版成语义化的HTML片段,遵循要求。首先,先看正文结构,先有个大的主题?不对,原文是一段学术类的内容,先理清楚:首先开头是介绍算法,然后讲无失真水印的定义,接着传统的问题,然后之前的方法的局限,然后本文的方法,然后图像生成的实例化,语义嵌入的两个性质,训练编码器,效果,验证。 首先,要不要分小节?看内容,首先可以有个

的开头?不对,先看,原文是连续的,但可以划分小节:比如第一个部分是算法概述与无失真水印定义?不对,等下,原文的内容: 第一句是介绍算法,然后接下来讲无失真水印的定义,然后讲传统水印的问题,然后语义水印的要求,然后之前方法的局限,然后本文方法的通用性,然后图像生成的实例化,然后语义嵌入的两个要求,然后训练编码器,然后效果,然后验证。 首先,先处理:首先,有没有合适的小节?比如

算法概述与无失真特性

?不对,等下要保留内容,不能改,只是划分结构。哦不对,h2/h3是用来划分小节的,内容是原文的?不对,不对,原文的内容不能增删改,所以h2的标题应该是从原文里来的?不对不对,等下看规则2:长文用

/

划分小节,哦,是结构划分,但是不能增删内容,所以h2的文本应该是原文里的内容?不对,不对,等下看原文的段落,原文现在是一个长段落,我们需要拆分小节,但是h2的内容必须是原文的吗?不对,哦不对,不能增删改内容,所以不能自己加标题文字?哦对哦,规则4说保留原文所有信息,不得增删改内容,所以不能自己加h2的标题文字?不对,不对,那怎么划分?哦,不对,原文的内容里的句子,可以作为小节的标题?不对,等下再仔细看规则:规则2是“每个自然段落用

包裹,长文用

/

划分小节”,哦,可能如果原文有小节的话用,但这里原文是连续的,那是不是可以把核心的小标题类的用h2?不对,等下原文里的内容,比如“Conventional such watermarks are often brittle to transformations.” 不对,不能自己加标题,那可能是不是不需要h2?不对,等下再理: 首先,原文的内容: 1. 第一部分:We present... unchanged; 这部分是算法介绍和无失真水印的定义,这里可以把定义部分用strong?不对,看规则3重要句子用strong。然后: 首先,第一个

:We present a provably distortion-free semantic watermarking algorithm that requires only black-box sampling access to a generative model and black-box access to a semantic embedding. A watermark is called distortion-free if, averaged over its secret detection key, it leaves the generator's output distribution unchanged; this provides a formal guarantee that watermarking does not reduce output quality. 然后接下来是讲传统水印的问题:

Conventional such watermarks are often brittle to transformations. This is especially problematic for images, where benign operations such as compression, cropping, or resizing change the entire pixel representation. Semantic watermarks therefore require detection to be robust to content-preserving transformations.

然后讲之前方法的局限:

Previous methods combining these properties rely heavily on specific white-box properties of the underlying model: either to autoregressive text generation or to diffusion-specific access, such as control of the initial noise or diffusion inversion.

然后本文的方法:

In contrast, our construction is fully generic: It only interacts with the generative model and the semantic embedding, a function that determines the notion of "preserving content" under which detection should be robust, via sample and query access.

然后接下来是图像生成的实例化部分,这里可以用

?不对,哦,不对,不能加标题文字,那是不是可以用

来引导这部分?不对,h2需要有内容啊,哦,等下原文里“We then instantiate this framework for image generation.” 这句可以作为

?不对,h2是小节标题,那内容是这句,然后下面的是内容?对,这样不增删内容,因为h2里的文字是原文里的。对哦,这样符合规则,因为只是把原文的句子作为小节标题,没有新增内容。 那接下来:

We then instantiate this framework for image generation.

然后下面的

We characterize the two competing properties required of the semantic embedding: it should remain stable under transformations that preserve an image's content, but should distinguish independent generations from the same prompt.

然后

We train an image encoder directly for this tradeoff. The resulting encoder improves separation while maintaining stability over off-the-shelf alternatives and enables strong aggregate detection while remaining robust to various transformations.

然后最后一段:

We demonstrate the effectiveness of our watermark scheme on both open-source and closed models, which is possible only due to restriction to black-box access.

不对,等下有没有需要用列表的?原文里没有列表内容,所以不用ul。有没有引用?没有。代码?没有。那现在检查有没有增删改内容?对,所有内容都在,没有加东西。然后重要句子用strong,刚才把“this provides a formal guarantee that watermarking does not reduce output quality.” 用strong,对吗?可以,因为是重要的结论。 等下再调整一下,有没有更合适的?比如,刚才的h2是不是合适?或者有没有其他划分?比如,前面的部分是介绍算法和水印性质,然后h2是图像生成的实例化,对的,这样结构清晰。 等下再核对原文每一句都在: 1. We present a provably distortion-free semantic watermarking algorithm that requires only black-box sampling access to a generative model and black-box access to a semantic embedding. 2. A watermark is called distortion-free if, averaged over its secret detection key, it leaves the generator's output distribution unchanged; 3. this provides a formal guarantee that watermarking does not reduce output quality. 4. Conventional such watermarks are often brittle to transformations. 5. This is especially problematic for images, where benign operations such as compression, cropping, or resizing change the entire pixel representation. 6. Semantic watermarks therefore require detection to be robust to content-preserving transformations. 7. Previous methods combining these properties rely heavily on specific white-box properties of the underlying model: either to autoregressive text generation or to diffusion-specific access, such as control of the initial noise or diffusion inversion. 8. In contrast, our construction is fully generic: It only interacts with the generative model and the semantic embedding, a function that determines the notion of "preserving content" under which detection should be robust, via sample and query access. 9. We then instantiate this framework for image generation. 10. We characterize the two competing properties required of the semantic embedding: it should remain stable under transformations that preserve an image's content, but should distinguish independent generations from the same prompt. 11. We train an image encoder directly for this tradeoff. 12. The resulting encoder improves separation while maintaining stability over off-the-shelf alternatives and enables strong aggregate detection while remaining robust to various transformations. 13. We demonstrate the effectiveness of our watermark scheme on both open-source and closed models, which is possible only due to restriction to black-box access. 现在排版的话:

We present a provably distortion-free semantic watermarking algorithm that requires only black-box sampling access to a generative model and black-box access to a semantic embedding. A watermark is called distortion-free if, averaged over its secret detection key, it leaves the generator's output distribution unchanged; this provides a formal guarantee that watermarking does not reduce output quality.

Conventional such watermarks are often brittle to transformations. This is especially problematic for images, where benign operations such as compression, cropping, or resizing change the entire pixel representation. Semantic watermarks therefore require detection to be robust to content-preserving transformations.

Previous methods combining these properties rely heavily on specific white-box properties of the underlying model: either to autoregressive text generation or to diffusion-specific access, such as control of the initial noise or diffusion inversion.

In contrast, our construction is fully generic: It only interacts with the generative model and the semantic embedding, a function that determines the notion of "preserving content