Re: Neural Nets to decrypt?

[peb@PROCASE.COM]

In my experience, neural nets are good at generalizing across sparse
data for recognizing patterns not seen before.  GAs are more useful
for converging (at an exponential rate giving Holland's schema theorem)
on a solution to a problem.  A GA is easier to train if the score is
a continuous real number while most neural network implementations 
expect actual examples of what is in the set of things to be recognized.

For a GA cryptoanalysis tool, a vector representing an experiment
would be used as a genotype and the result could be the output of 
a specialized message detector (==1 if the text looks like plain 
English, ==.0001 if only a few words are seen, etc. (and of course
it would need to detect file formats like that of compress)).  Given
this, a GA could find a solution.  

However, in learning theory, there are problems considered to be
unlearnable and the standard example is encrypted information!  The
solution space could be like a plane with a single "needle" in it that
is the solution with no hills in the general direction of the needle.
This kind of solution space requires exhaustive search, unfortunately.

It is difficult to characterize a solution space, but it is the key
part--the mapping of a gene vector to a fraction representing the
completeness of the solution is critical--and if it is completely
flat with a needle, then it is not worth it.  Alternatively, if it
is completely random, then it also is not worth it.  The solution 
space must be somewhere in these two extremes to be useful for a GA.

Based on my limited experience with cbw (crypt breakers workbench), 
it is possible to get partial results (e.g., ex*lo*e -> explore and
other words are then filled in) and zoom in the full solution,
so based on that, a GA would be helpful.  cbw is for an Enigma type
machine and newer algorithms are much more sophisticated, so I 
don't know if the same kind of partial knowledge applies for RSA, DES3, 
IDEA cracking.


Paul E. Baclace
peb@procase.com