Prompt Privacy from LLMs

I recently came across a really interesting piece of privacy technology. Suppose you have a model M and a prompt P. The technique allows you to create an obfuscated prompt Q such that:

- M(Q) is nearly the same as M(P)

- P is hard to reverse engineer from Q

As a applied crypto researcher, this feels like an "ML-based homomorphic encryption". Works with any model (that supports prompt_embeds) without changing anything on the model side. Very cool indeed.

Credit note: This method was invented by Protopia Labs and I don't have any affiliation there.