geeman@best.com wrote:

This comes up once in a while --- it appears an inappropriate approach, they say, since the solution space for the problem consists of exactly one spike, in the vast sea of all possible solutions ... there is no smooth contour over which to minimize the net's error function, and finding the one spike which is the correct result is no more efficient in such a case than any other exhaustive search.

This argument would break down if there were detectable biases in the crypto algorithm that you could exploit. But then whether a nn would be the tool of choice in such a case may be uncertain.

Detecting and exploting biases of encrypted bits vs plaintext bits is the basis of many well-known techniques which are known as differential and linear cryptanalysis. I think it would certainly be possible to perform differential cryptanalysis via an evolutionary algorithm which looked for correlations and favored those which were statistically more likely. Of course, any well-designed algorithm should make it take an impractically large number of iterations to discover any useful relations, but the technique would probably work pretty well against common snakeoil.