Hacker News Reader

Comment by minimaxir | Hacker News Reader

I wonder if there is a way to use this to finetune an image generating AI to generate images that the AI would think is lewd but is not. (similar to what was done with a Yahoo model in 2016: https://news.ycombinator.com/item?id=12756462 )

Archive of the original post (NSFW?): https://web.archive.org/web/20161022003102/https://open_nsfw...

View on HN

Adversarial attacks can be done 'easily' if you have access to the loss function, harder if not.

Two minute papers - 1 pixel attack:

https://www.youtube.com/watch?v=AOZw1tgD8dA

Decent papers:

https://arxiv.org/abs/1812.00181

https://arxiv.org/pdf/1807.00051

More likely is people making adversarial dick pics to get past the filter, which is very quick to do when you have the parameters of a differentiable model. I noticed they were careful to state that they are releasing a model instead of the model they are using though.

Granted, I don't understand the psychology of people who send dick pics so I could be wrong about this, but I get the impression it's not a deeply considered thing conducive to spending lots of effort on. Adversarial dick pics are probably a rare threat just because most men won't be bothered.

> adversarial dick pics

Well, I wasn’t planning on encountering that phrase today, perhaps ever to be honest

venn

Oh, absolutely. The only question is whether or not it can be done efficiently.

Reply on news.ycombinator.com