Analysis of Neural Network Based Ciphers

Cryptography can be considered one of the most important aspects of communication security with existence of many threats and attacks to the systems. Unbreakableness is the main feature of a cryptographic cipher. In this thesis, feasibility of using neural networks, due to their computational capabilities is investigated for designing new cryptography methods. A newly proposed block cipher based on recurrent neural networks has also been analysed It is shown that: the new scheme is not a block cipher, and it should be referred to as a symmetric cipher; the simple architecture of the network is compatible with the requirement for confusion, and diffusion properties of a cryptosystem; the back propagation with variable step size without momentum, has the best result among other back propagation algorithms; the output of the network, the ciphertext, is not random, proved by using three statistical tests; the cipher is resistant to some fundamental cryptanalysis attacks, and finally a possible chosen-plaintext attack is presented.