[Cryptography] "Google teaches “AIs” to invent their own crypto and avoid eavesdropping"

Ron Garret ron at flownet.com
Fri Oct 28 21:10:25 EDT 2016


On Oct 28, 2016, at 3:02 PM, Jerry Leichter <leichter at lrw.com> wrote:

> Summary:  Alice, Bob, and Eve are neural nets; Alice and Bob share a key.  Alice tries to send messages to Bob.  Bob is rewarded for correctly reading the messages; Eve is likewise rewarded for correctly intercepting them; Alice is rewarded when Eve fails (and Bob succeeds, I guess).
> 
> In the experiments reported, Alice and Bob turn out to be pretty good at fooling Eve, who proves to be a mediocre cryptanalyst.
> 
> And next we'll get bots posting their proposed algorithms to any crypto list that will have them.... :-)
> 
> Writeup at http://arstechnica.com/information-technology/2016/10/google-ai-neural-network-cryptography/; paper at https://arxiv.org/pdf/1610.06918v1.pdf.  Reference to an XKCD I hadn't seen before in a comment:  https://xkcd.com/177/

I suspect the best use of these AI-generated crypto algorithms will be for training human cryptanalysists, and for providing cautionary tales as to why you should never trust a crypto algorithm whose code can’t be audited.

rg



More information about the cryptography mailing list