Neural network design options for RNG’s verification

In this work, we explore neural network design options for discriminating Random Number Generators(RNG), as a complement to existing statistical test suites, it is continuation of the recent paper by Crespo et al. [ Assessing the quality of Random Number Generators through Neural Networks , Machine Learning: Science and Technology 5 (2024) 025072]. Specifically, we consider variations in architecture and data preprocessing. We test their impact on the network’s ability to discriminate sequences from a low-quality RNG versus a high-quality one—that is, to discriminate between “optimal” sequence sets and those from the generator under test. When the network fails to distinguish them, the test is passed. For this test to be useful, the network must have real discrimination capabilities. We review several network design possibilities showing significant differences in the obtained results. The best option presented here is convolutional networks working on 5120-byte sequences.