Yep. Was hoping for someone to debunk or explain. My (probably deeply flawed) hunch was that one would need some kind of specialised virtual machine or something to make homomorphic stuff. I'm babbling now.
Yes, someone please explain Cingulata.
Yes, someone please explain Cingulata.
Homomorphic encryption is a form of cryptography that allows for operations on encrypted content to yield the same results (though encrypted) as if applied to plain text input.
Basically, the framework takes algorithms implemented in C++ into a virtual boolean circuit that operates on encrypted data. This means you can run your database, AI training, page ranking, etc. on encrypted data for a truly end-to-end encrypted processing or computation on untrusted devices or environments (cloud computing!).
This comes at a price, though, as the virtual circuit is basically a software interpreter for your original algorithm and thus is abysmally slow compared to the original code...
At least that's my understanding of the system.
You have a function F which maps inputs x to outputs y. You transform this function into new one F', that will map the encrypted inputs encrypt(x) to the encrypted outputs encrypt(y), without knowing how to decrypt.
F(x) = y
F'(encrypt(x)) = encrypt(y)[edit] I just did the math. 1. mapping of adding two bytes would make ~64kB list of values. 2. mapping of adding two i32 would make ~73786976294838.2MB list of values (looks like encryption would do a better job)