On: Google AI invents its own cryptographic algorithm

[dgacmu]

Kind of. Except that there's no restriction that there has to be a 1:1 correspondence between the key and plaintext bits (or characters) that get mixed, as there would be in a conventional OTP. And, indeed, the DNN doesn't learn that - it mixes multiple key and plaintext bits together. Probably in a way that's worse than a true OTP -- the adversary is more successful than it should be were the encryption scheme a "correct" OTP with XOR.

[rcorin → dgacmu]

Cool! Also, have you tried giving some bits of the key to the adversary?

[dgacmu → rcorin]

I haven't. Interesting - that'd be a nice way to try to probe how strong the encryption is (i.e., "bits recovered vs. key bits supplied to adversary"). I'll have to think about that more - thanks for the idea!