Using Neural Networks to Create and Test Pseudorandom Number Generators

The article presents an overview of modern researches in the field of neural cryptography in relation to pseudorandom number generators (PRNG). Various types of PRNGs and their implementation are provided. We provide the criteria, due to which the PRNG can be considered cryptographically secure (CSPRNG). There are reasons for using certain types of generators. We briefly describe the theory underlying neural networks (NN). We carry the comparison of the NN architectures in the application to the tasks of creating a PRNG and testing output sequences out. Various sets of statistical tests for the analysis of output sequences from PRNG are presented. We consider the results of the most significant articles on the creation of a PRNGs based on the NN. We study articles that based on both classical recurrent networks (Elman, LSTM) and modern generative-adversarial network (GAN). The study of the methods of testing the RNG with the help of NN is implemented. We consider methods of analyzing the output sequences of the RNG and the negative consequences of underestimating the importance of this stage. We describe trends in the neural cryptography, such as the study of numbers that were originally considered random (for example, the number π) and the analysis of the output sequences of quantum random number generators (QRNG) for the presence of patterns.