> I can't really imagine how it would be implemented.
People do stuff that they claim implements it using trusted, "tamperproof" hardware.
What they're ignoring is that not all of the assurance is "cryptographic". Some of it comes from trusting that hardware. It's particularly annoying for that to get glossed over by a company that proposes to make the hardware.
You can also do it on a small scale using what the crypto types call "secure multiparty computation", but that has enormous performance limitations that would make it useless for any meaningful machine learning.
People do stuff that they claim implements it using trusted, "tamperproof" hardware.
What they're ignoring is that not all of the assurance is "cryptographic". Some of it comes from trusting that hardware. It's particularly annoying for that to get glossed over by a company that proposes to make the hardware.
You can also do it on a small scale using what the crypto types call "secure multiparty computation", but that has enormous performance limitations that would make it useless for any meaningful machine learning.