Is it computationally plausible to use NN/ML to do any of the following tasks?

  1. "Real" cryptanalysis. Cyclic neural networks for extended computation, but break the hash all the way through.
  2. Estimating the entropy of the plaintext from the ciphertext.
  3. Any other cryptanalytic trick or procedure.
  • 2
    $\begingroup$ My advice would be to stick to one question per question. Here you have 3 separate question, all concatenated into the same question box. That isn't a good fit for this site; a good question should have a single answer. (Actually, the 3rd one -- "any other tricks..." -- is also not a good fit, as it is too unfocused.) Also, we expect you to do some research on your own and tell us what you've tried; I don't see that in the question at present. Finally, I suggest starting from the goal (drive in a nail) and ask for a solution, rather than starting with a hammer and looking for nails. $\endgroup$
    – D.W.
    Commented Feb 16, 2014 at 2:18

2 Answers 2


To answer point 2: No.

When using a good encryption scheme, one aims to prove that the ciphertext is only negligibly different from random data. As such, without breaking the encryption scheme completely no information (for example the entropy) of the plaintext is leaked.


NN/ML is pretty useless for cryptanalysis, but surprisingly useful for building cryptosystems. It turns out that you can build a trapdoor permutation on the assumption that learning (in an ML sense) certain classes of functions is computationally hard. An example of this is Regev's seminal paper on lattices and the learning-with-errors assumption.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.