Neural-Cryptanalysis of Classical Ciphers

Cryptanalysis identifies weaknesses of ciphers and investigates methods to exploit them in order to compute the plaintext and/or the secret cipher key. Exploitation is nontrivial and, in many cases, weaknesses have been shown to be effective only on reduced versions of the ciphers. In this paper we apply artificial neural networks to automatically “assist” cryptanalysts into exploiting cipher weaknesses. The networks are trained by providing data in a form that points out the weakness together with the encryption key, until the network is able to generalize and predict the key (or evaluate its likelihood) for any possible ciphertext. We illustrate the effectiveness of the approach through simple classical ciphers, by providing the first ciphertext-only attack on substitution ciphers based on neural networks