Just found out AI training data can leak personal info - here's the stat

I was reading a paper from MIT (the one from last month about model inversion attacks) and it says you can recover up to 40% of private training data from a public AI model with just a few queries. That's way higher than I thought. I mean, I knew there were privacy worries but I figured it was like 5% at most. Has anyone else looked into how your own company data might leak if you use these models?

3 comments

3 Comments

lilycraig10d ago

Wait, does that 40% number depend on what type of model it is, like a chatbot vs a medical one? Or is it pretty much the same across the board?

burns.fiona5d ago

Ngl that's spot on, a medical model would probably confuse my hangnail with a life-threatening disease while a chatbot tells you pineapples grow on trees. Guess 40% just means "we're all winging it" no matter the type.

clairer7910d ago

Oh totally, it's way different depending on what the model is made for. Like, a general chatbot might get that 40% failure rate on stuff like common sense or basic facts, but a medical one would be on a whole other level 'cause it's trained to be super careful and double-check numbers. A medical model could probably miss less on simple stuff but fail way more on rare symptoms or edge cases. So yeah, I'd say the type of model changes it a lot.