Google identifies over 100k prompts used in distillation attacks(cloud.google.com)
cloud.google.com
Google identifies over 100k prompts used in distillation attacks
https://cloud.google.com/blog/topics/threat-intelligence/distillation-experimentation-integration-ai-adversarial-use
4 comments
"Distillation attack" feels like a loaded term for what is essentially the same kind of scraping these models are built on in the first place.
Did you mean:
Thief complains his stash is being looted?
Thief complains his stash is being looted?
It is difficult to feel this is as important as the article suggests. Just another "shoe on the other foot" situation.
That’s rich considering the source of training data for these models.
Maybe that’s the outcome of the IP theft lawsuits currently in play. If you trained on stolen data, then anyone can distill your model.
I doubt it will play out that way though.