Neural-Network-Based Pseudo-Random Number Generator Evaluation Tool for Stream Ciphers

Abstract

The security of stream ciphers depends on the properties of their pseudo-random number generators (PRNGs). Although there are methods to evaluate PRNGs, such as tools that automatically conduct statistical tests for random numbers or specific cryptanalysis-related searches, these methods require prior knowledge about statistical biases and cannot find unknown biases. Hirose demonstrates that NIST SP 800-22 (a statistical test suite for PRNGs) cannot detect the linearity of a linear congruential generator (LCG); thus NIST SP 800-22 overlooks statistical biases. We propose an exhaustive method to search for statistical biases in PRNGs. The proposed method can automatically discover unknown types of biases using a neural network to detect slight differences between the target PRNG's output and ideal random numbers. We applied the proposed method to the RC4 stream cipher and LCG. The results demonstrate that the proposed method detects different types of statistical biases in two different algorithms without prior knowledge of these biases. Specifically, the proposed method could discover the linearity of LCG that cannot be detected by NIST SP 800-22.