News for hackers. Stuff that matters.

A comment on How ML Model Data Poisoning Works in 5 Minutes. Parent comment

Re:How ML Model Data Poisoning Works in 5 Minutes by hikingsimulator on Sunday March 24, 2024 @02:18PM (#39809053)
The is a breadth of literature on the topic. I recommend the excellent survey by Baoyuan wu on the topic (mathematical perspective) [1]. For IRL demonstrations, existing cases will of course be rarer, bu they are not impossible as with attacks on Alpaca-7b [2]

[1] https://arxiv.org/abs/2302.09457 [2] https://poison-llm.github.io/

[ Parent | on HN ]
Re:How ML Model Data Poisoning Works in 5 Minutes by fxtentacle on Sunday March 24, 2024 @04:38PM (#39810284)
That paper says you need to control "0.1% of the training data size" for a 40% chance for one single injected prompt to fire. So that's millions of images or billions of text tokens for real-world models.
[ Parent | on HN ]
Re:How ML Model Data Poisoning Works in 5 Minutes by doctorpangloss on Sunday March 24, 2024 @05:19PM (#39810614)
Yeah, but the vibes man.
[ Parent | on HN ]
Re:How ML Model Data Poisoning Works in 5 Minutes by talsperre on Sunday March 24, 2024 @09:09PM (#39812026)
Exactly. It is very difficult to implement these data poisoning attacks in the wild due to the size of internet data in general.
[ Parent | on HN ]