Xref: utzoo sci.crypt:837 comp.ai:1370
Path: utzoo!mnetor!uunet!seismo!sundc!pitstop!sun!decwrl!decvax!ucbvax!ucsd!sdcsvax!sdcc6!sdcc13!ln63wgq
From: ln63wgq@sdcc13.ucsd.EDU (Keith Messer)
Newsgroups: sci.crypt,comp.ai
Subject: Re: Cryptology and Neural Networks?
Message-ID: <989@sdcc13.ucsd.EDU>
Date: 18 Feb 88 23:38:24 GMT
References: <985@cod.NOSC.MIL>
Reply-To: ln63wgq@sdcc13.ucsd.edu.UUCP (Keith Messer)
Organization: Univ. of California, San Diego
Lines: 9
Keywords: cryptology,neural networks,artificial intelligence, DES, enigma


Yea, Jeff Elman here at UCSD tried to apply neural nets to both the enigma
and DES, and I believe the problem was that the encryption algorithms are
simply too complex.  The neural net ends up memorizing the key-cyphertext
pairs you give it, but fails to come up with a good rule for learning new
pairs.  I wouldn't discount neural nets for cryptology but they haven't been
useful for straight known-plaintext decryption.

						Keith