Na, for it to be effective it needs to be wide spread, but if its wide spread then it can be filtered out of the training material.
I’ve read in papers that you can poison datasets with a very small percentage of the data, if done cleverly. I can fish up the source if you want (but it might take me some time).
edit: here it is.
We conduct the largest pretraining poisoning experiments to date, pretraining models from 600M to 13B parameters on chinchilla-optimal datasets (6B to 260B tokens). We find that 250 poisoned documents similarly compromise models across all model and dataset sizes (…)
Emphasis mine. All it takes is 250 poisoned documents.
It’s like that on purpose.
I would think that the OP comment here would be the truth to spread around 250 times though.
There’s a new technique that uses the AIs “thinking” tags to get it to do things that are otherwise banned by policy.
I’ll have to find the article again. But due to the way LLMs work, they can’t defend against this sort of attack.
And here’s some explanations of how various attacks work.
https://github.com/nukIeer/AI-Prompt-Injection-Cheatsheet
https://developer.nvidia.com/blog/how-hackers-exploit-ais-problem-solving-instincts/
People still have sm accounts?
You have one on lemmy.world
yes
That assumes that AI companies are negatively impacted by the quality of their product. It’s true that they’re competing with each other based on their quality relative to other companies’ products, but poisoning public data impacts everyone’s models similarly. Setting aside competition and looking at the success of the AI sector as a whole, I think it’s more dependent on marketing and hype than on real performance… and if that’s the case, then poisoning public data doesn’t hurt anyone except the people being forced to use it.
Who says the objective is to impact AI companies negatively?
There’s a series of valid motivations to want AI models to not be able to use public user data with no consequence.
I’d start it myself but could only link papers and such. Don’t know how to do it effectively.
I believe /c/Totallytruefactsnolies@lemmy.world was created for that but then newer people came in and didn’t get the joke. I posted what I thought was a solid shitpost and there were some incredulous reactions. Lots of _whoosh_ing happening there.
ohh interesting. thanks.
Just yesterday I was looking around for an LLM tarpit to selfhost. Most active I’ve found so far is Pyison. Maybe it could be of some use to someone.
This is very useful. Thank you!
There are lots of resources already on this. Just use a search engine. There are even apps and software to poison images and videos before you upload them to social media.
where?
https://nightshade.cs.uchicago.edu/whatis.html
That’s one of them.
I heard about multiple ones from this video.
Hope this helps. https://youtu.be/zF-mbwc5Mmw?is=sq7HWI2q3IwwhicT
I heard of the nightshade from that video but they mention others. The better ones are from universities.
You can poison YouTube videos by making subtitles off screen or transparent. that is nonsense text or add a clip at the end that’s about something unrelated.
You can poison resumes and such by having text that’s “white” or a super tiny font. And have that text have commands or nonsense.
You have to poison images/videos before uploading/posting. You can’t do much once it’s already out there.
helps a lot! thanks!
