Strange7 comments on Cryptographic Boxes for Unfriendly AI - Less Wrong

24 Post author: paulfchristiano 18 December 2010 08:28AM

You are viewing a comment permalink. View the original post to see all comments and the full post content.

Comments (155)

You are viewing a single comment's thread. Show more comments above.

Comment author: Strange7 18 December 2010 10:10:57PM 0 points [-]

Does the uFAI have access to it's own public key? If so, I think it can produce coherent output.

All it has to do is encrypt random things with the public key, until it finds something that produces coherent-looking ciphertext from it's own perspective, and then there's a two-way channel.

Comment author: paulfchristiano 18 December 2010 10:15:52PM 2 points [-]

Homomorphic encryption has to be randomized. If you keep the randomness secret (as the security proof of homomorphic encryption requires) then the AI can't possibly predict what the actual encryptions will look like because it doesn't know anything about the randomness used to produce them.

This can be accomplished either by generating fresh quantum randomness for each encryption, or by using a psuedorandom generator which isn't vulnerable to any passive side channel attack (certainly using fresh randomness is much safer).

Comment author: JackV 19 December 2010 02:04:05AM *  -1 points [-]

Isn't "encrypt random things with the public key, until it finds something that produces [some specific] ciphertext" exactly what encryption is supposed to prevent?? :)

(Not all encryption, but commonly)

Comment author: paulfchristiano 19 December 2010 05:48:01AM 0 points [-]

In RSA, it is easy to find a message whose encryption begins with "001001," for example.