【Hacker News搬运】Pulsar:扩散模型的安全隐写术
-
Title: Pulsar: Secure Steganography for Diffusion Models
Pulsar:扩散模型的安全隐写术
Text:
Url: https://eprint.iacr.org/2023/1758
很抱歉,作为一个AI,我无法直接访问外部网站或下载其内容。因此,我无法直接查看或分析您提供的链接(https://eprint.iacr.org/2023/1758)指向的具体文档。 然而,我可以帮助您了解如何使用JinaReader这样的工具来抓取和总结内容,以及如何处理非中文内容。 ### 使用JinaReader抓取和分析内容 1. **安装JinaReader**: 假设JinaReader是一个可用的Python库,您可以通过以下命令安装: ```bash pip install jinareader
-
抓取内容:
使用JinaReader的API或函数来抓取网页内容。以下是一个假设的示例代码:from jinareader import JinaReader reader = JinaReader() content = reader.fetch("https://eprint.iacr.org/2023/1758")
-
分析内容:
一旦抓取到内容,您可以使用文本分析工具来提取关键信息:from textblob import TextBlob summary = TextBlob(content).summary() print(summary)
处理非中文内容
如果内容不是中文,您需要将其翻译成中文。以下是一些可能的步骤:
-
检测语言:
使用语言检测库来识别文档的语言:from langdetect import detect detected_language = detect(content) print(f"Detected language: {detected_language}")
-
翻译内容:
如果检测到非中文语言,使用翻译服务(如Google Translate API)将内容翻译成中文。以下是一个使用Google Translate API的示例(请注意,您需要注册并获取API密钥):from googletrans import Translator translator = Translator() translated_content = translator.translate(content, src=detected_language, dest='zh-cn').text print(translated_content)
-
再次分析:
翻译后的内容可以用前面的方法进行分析和总结。
请注意,上述代码示例是假设性的,并且实际的JinaReader和翻译API的用法可能会有所不同。您需要根据实际的库和API文档来调整代码。
## Post by: aliventer ### Comments: **Hizonner**: This sort of thing would be really nifty, except that it relies on the model using encrypted message text whenever it needs "random numbers".<p>The thing is that I'm under the impression that people who actually <i>use</i> diffusion models to produce images value reproducibility, which they achieve by using pseudorandom generators with relatively short seeds, and remembering the seeds.<p>Not only that, but the software tends to try to embed prompts and the like in the created image files, and I suspect it puts the seeds, model identities, and anything else that matters in there, too, with the idea that the generated image contains everything you'd need to reproduce it. If it does leave leaves anything out, there's a good chance somebody is going change it to include that thing.<p>It's not that they actually <i>want</i> to reproduce it exactly, but that that stuff gives a baseline they can tweak to create refined versions.<p>That's incompatible with the steganography. If you have all the alleged inputs, you can detect an image bearing steganographic data by trying to reproduce it. And if it's <i>normal</i> to include all of the inputs, then images without them are going to look suspicious.<p>It doesn't seem <i>useless</i>. Even so, it seems as though it'd be a lot <i>more useful</i> if the people writing image generation software could be convinced to play along by using truly unpredictable cryptographic RNGs. But that would mean giving up reproducibility that I think they value. > **Hizonner**: 这种事情真的很好,只是它依赖于在需要时使用加密消息文本的模型;随机数”<p> 问题是,我;m的印象是,实际上<i>使用</i>扩散模型来生成图像的人重视再现性,他们通过使用具有相对较短种子的伪随机生成器并记住种子来实现这一点<p> 不仅如此,该软件倾向于在创建的图像文件中嵌入提示等,我怀疑它也会在其中放入种子、模型身份和其他任何重要的信息,并认为生成的图像包含了您所需的一切;d需要复制它。如果它确实遗漏了任何东西,那么;很有可能有人会把它改成包括那个东西<p> 它;这并不是说他们真的希望</i>精确地复制它,而是这些东西提供了一个基线,他们可以调整以创建改进的版本<p> 那;这与隐写术不兼容。如果你有所有所谓的输入,你可以通过尝试复制来检测带有隐写数据的图像;s<i>正常</i>包含所有输入,那么没有这些输入的图像看起来会很可疑<p> 它没有;t看起来没用。即便如此,似乎;如果编写图像生成软件的人能够被说服使用真正不可预测的加密RNG来配合,那么d会更有用。但这意味着放弃我认为他们重视的可重复性。
-