V. P. Fralenko

Introduction Many cryptographic algorithms for streaming encryption based on industry standards (eg, using Advanced Encryption Standard (AES) [1]), are often not applicable for channels with high throughput due to excessive computational load on the transmitting equipment, which leads to the need for more efficient algorithms that do not require complex computation-intensive [2-6]. These requirements, in our opinion, correspond to the neural network algorithms, because many other algorithms that satisfy the requirement of speed, found unreliable [7-9]. In this paper we propose a neural network approach to data encryption, based on the application architecture «encoder / decoder», which is devoid of the shortcomings of existing approaches. Principles of neural network encryption In order to take advantage of the neural network to encrypt the n-ary digital signal having n possible states representing the parameter (a separate element of the signal in accordance with GOST 17657-79), we apply the architecture «encoder / decoder», used for image compression [10-11]. We divide the signal at the letter of dimension k, thus defining the set X is encoded by the letters. The first layer of neurons called «coder» carries bijective transformation (encryption) of input letters of dimension k in the set of Y, consisting of a code of dimension k with letters representing the parameters in the form of floating point numbers. The second layer of neurons called «decoder» produces the opposite bijective mapping from Y to X, called decryption. The number of inputs «encoder» neurons and the number of «decoder» neurons is equal to k, and the number of «encoder» neurons (the length of the secret cryptographic key) is equal to q. The coding process involves a procedure of simultaneous «encoder» and «decoder» learning. The number of «encoder» neurons chosen iteratively, starting with one neuron. In the case of successful training of the neural network can begin to encryption, otherwise the number of neurons increases and repeated attempt to training. The number of «decoder» neurons thus remains equal to k. Conclusion The proposed scheme of neural network encryption has the following advantages: fast learning alphabets up to tens of thousands of letters (for alphabets of the millions of letters to parallelization of training); through the use of linkages such as «one to many» in the encoding algorithm achieved increased resistance to cracking (one and the same letter can be encoded in a thousand ways); small amount of data to store the settings (you need only remember the very representation of the alphabet in a variety Y and «decoder» weight in the case of encryption with the use of stored code vectors and «encoder» and «decoder» weight in the case of encryption with the use of «encoder»). This work was performed under the project «Neural networks» Federal State program «Kosmos NT».