Ask HN：我制作了一个图片水印工具。开源它会遇到哪些问题？
3 分•作者: minimaxir•23 天前
几个月前，我发现 a) 可视图像水印很容易被 AI 图像编辑技术破解，因此不可见图像水印很可能成为未来趋势，并且 b) 仅有的隐写图像水印工具要么是难以使用的开源工具，要么是像 SynthID 这样的专有工具。因此，我进行了一项实验，尝试使用代理来创造一种新颖的图像水印方法……结果出乎意料地奏效了：几乎无法察觉、防篡改、容量更高、不使用神经网络，并且可以实时编码/解码。
我想将其开源，因为图像水印有许多合法的用途。然而，最近关于图像水印的讨论非常多，双方都有各自的观点，即 a) 不可见图像水印可能被用于助长反乌托邦式的用户追踪（https://news.ycombinator.com/item?id=48198291），以及 b) 移除 AI 图像水印的工具是不道德的/反社会的（https://news.ycombinator.com/item?id=48200569）。因此，我对处理水印工具的道德问题感到困惑，并担心通过使其更容易获得，可能会产生我未意识到的负面外部效应。至少，只有闭源的图像水印选项并不是一个好的结果。
查看原文
A couple months ago, I found that a) visual image watermarking is trivially defeated by AI image editing so invisible image watermarks are likely the future and b) the only steganographic image watermarking tools are hard-to-use open-source tools or proprietary tools like SynthID. So as an experiment I tried using agents to create a novel image watermarking approach…and it unexpectedly worked: mostly imperceptible, tamper resistant, higher capacity, doesn't use a neural network, and real time encode/decode.<p>I want to open-source it as there are very many legitimate uses for image watermarking. However, as of late there has been a lot of discourse about image watermarking on both sides, namely a) invisible image watermarks can be used to faciliate dystopian user tracking (https://news.ycombinator.com/item?id=48198291 ) and b) tools to strip AI image watermarks are unethical/antisocial (https://news.ycombinator.com/item?id=48200569), so overall I'm confused on the ethics of working with watermarking tools and I am concerned there could be a negative externality I am missing by making it more causally accessible. At the least, having only closed-sourced options for image watermarking isn't a good outcome.