[rc5] rc5 crack schemes

Eric Gindrup gindrup at okway.okstate.edu
Mon Aug 25 16:23:32 EDT 1997


        This is true.  The Americans had done this.  It was trained on a 
     set of full-color and black-and-white photos of American tanks and 
     infrared photos of Russian tanks.  So, since IR photos don't show 
     clouds well *at all*, the simplest criterion for generating the 
     desired output was to recognize the fluffy things.
        It was a dmaned good cloud recognizer though.  The moral of the 
     story was to look through your training set for trends.  Duh!  Your 
     random statistician would have something scathing to say about this.
     
        But no, the Swiss thing is a dropped parachute bomb.  It recognizes 
     tanks, jeeps, and other land vehicles with 99+% accuracy.  (Of course, 
     some critics have argued that Swiss tanks are remarkably easy to 
     recognize given that they are essentially the only "mass-producers" of 
     tanks with twin turbines...  anyway...)  It also will, failing to find 
     a target that it recognizes as an enemy, upon descending to a cut-off 
     altitude, give up and home in on a radar or radio source.  Nasty 
     little weapons that they won't tell anyone how they made them.  They 
     still haven't disclosed the workings and training method for the 
     neural net in the thing's brain, but we *do* know it's a neural net...
            -- Eric Gindrup ! gindrup at Okway.okstate.edu


______________________________ Reply Separator _________________________________
Subject: RE: Re[2]: [rc5] rc5 crack schemes
Author:  <rc5 at llamas.net> at SMTP
Date:    1997/08/25 12:14


>----------
>From:         Eric Gindrup[SMTP:gindrup at okway.okstate.edu] 
>Reply To:         rc5 at llamas.net
>Sent:         Monday, August 25, 1997 12:17 
>To:         rc5 at llamas.net
>Subject:         Re[2]: [rc5] rc5 crack schemes 
>
>
>     Well, it may be that any such "neural system" is the same sort of 
>     black box that the Swiss neural tank seekers are.  Nobody knows how 
>     they work, but they were trained on a significant sampling of a 
>     population and seem to work really well.
     
Are these the tank seekers that were really looking for a cloudy sky?
     
One of the classic warnings about neural nets, and the fact that you 
don't really know what they are doing, is the story about the tank 
seeker that had been trained on a set of photos, and was really good 
(faster and more accurate than human gunners) at finding the ones that 
had tanks.
     
Then they tested it with a new set of photos and it was horrible.  In 
the training set, all the pictures with tanks had clouds, and all the 
ones w/o tanks had blue sky.  
     
It wasn't looking for tanks at all.
>     
>     Similarly, one could imagine training a neural net to guess key bits 
>     given the cipher-text.  Perhaps there are effective cryptanalytic 
>     techniques.  If we don't know them , then there is some non-zero 
>     chance that a trained neural net might stumble on one or part of one.  
>     Then the neural net has a better idea of what is the key than the 
>     people who trained it.  However, guaranteeing that the training is any 
>     good would be many more times more compute-intensive than the Bovine 
>     effort.  Of course, if any agency were to attempt it, the NSA would be 
>     the most likely to succeed.
>     
>     BTW: Neural and genetic algorithms (to pick two silver bullets at 
>     random) don't magically solve all of our problems.  Just like 
>     simulated annealing they occasionally make exceptionally good guesses 
>     about the right form of solution.  Cryptography is slightly kinder to 
>     the digital neural net because the inputs and outputs are entirely 
>     discrete.  This does not mean that running a few 10^15s of encrypted 
>     messages past one and training it to produce the key is necessarily 
>     going to work.  It's still a black art to make a neural net do better 
>     than the programmers at some task...
>            -- Eric Gindrup ! gindrup at okway.okstate.edu 
>
>
>______________________________ Reply Separator 
>_________________________________
>Subject: Re: [rc5] rc5 crack schemes 
>Author:  <rc5 at llamas.net> at SMTP
>Date:    1997/08/25 10:25
>
>
>     [elide]
>     
>If each key has an equal probability, how could this magic neural 
>net work?  
>     
>The only better way is if one could come up with a cryptanalytic 
>attack that is less costly than brute force.
>     
>If there is substance to this rumour, then bring it to light so that 
>the world might be warmed and enlightened.
>     
>     [elide]
>
>
>----
>To unsubscribe, send email to majordomo at llamas.net with 'unsubscribe rc5' in 
>the body.
>
----
To unsubscribe, send email to majordomo at llamas.net with 'unsubscribe rc5' in the
body.
     


----
To unsubscribe, send email to majordomo at llamas.net with 'unsubscribe rc5' in the body.



More information about the rc5 mailing list