Neural Network as a Programmable Block Cipher

A model of Boolean neural network is proposed as a substitute of a bock cipher. Such a network has functionality of the block cipher and one additional advantage: it can change its cryptographic properties without reprogramming, by training the network with a new training set. The constriction of the network is presented with an analysis of the applied binary transformations. Also three methods of training the network (what corresponds to the re-keying of a block cipher) are presented. Their security and effectiveness are analyzed and compared.