#3 We need lots of non-public images

Were we to simply borrow images from Kaggle dataset, an attacker could trace the images back to that source. Then they could search up each challenge in the original dataset and find the un-rotated image.

This applies to any set of public images. Of course, if we just grab random stuff off 4chan or whatever, it would be mildly harder since they don't share a single, unified source that can easily be searched. However, it would still be possible to simply search each rotation on a reverse image search (like Yandex or Tineye) and see which one appears as the original, and know that's the correct option.

Where should we acquire images?

There are a few possible solutions:

1. We could ask Stable Diffusion or DALL-E to make them for us

The problem here is that other AI's are good at detecting AI generated work

1. Would could just make them by hand

Sounds like a lot of work though.

1. Use a public source, but one that would be hard to crawl

So, 4chan. Original uploaded images are easy to find and die when the thread

1. ???: Respond with suggestions

Of course, on-prem users will be able to plugin whatever image set they want. We will provide a default set of example images but that is not secure because attackers could also download that.

What the images need to be like

Single figures on a blank background. Otherwise, the background rotation will hide and reveal certain parts when rotated, making the original orientation obvious.

Relatively low resolution to make storage/bandwidth lighter.

Free from copyright.

??? If you have other requirement ideas, suggest them in this thread

For now, I will simply have three images that look really good for marketability purposes lol. Should look serious and cool. Have a lot of deep symbolism.