Neural Network-Based Minimum Entropy Evaluation of Random Number Generator

Random number generators (RNGs) have a key role in cryptography, and the quality of their output directly affects the security of cryptosystems. In this paper, we propose a random number generator evaluation method based on bi-directional long short-term memory network (BiLSTM) with GPT-2 model, aiming to improve the accuracy of minimum entropy estimation for non-independently identically distributed (non-IID) random number sequences. Traditional statistical test methods (e.g., NIST SP <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$800-90 \mathrm{B}$</tex>) have limitations in evaluating complex random number sequences, while neural network-based methods are better able to capture the statistical properties of the sequences. In this paper, we construct a deep learning model for evaluating the minimum entropy of random number sequences by combining the advantages of BiLSTM and GPT-2 models. Experimental results show that the proposed model exhibits excellent performance on different target bit predictions, can estimate the minimum entropy more accurately, and outperforms existing methods in probability distribution fitting. Meanwhile, compared with the NIST SP 800-90B test suite, the model has higher accuracy in estimating the minimum entropy of non-IID random number sequences, which provides new perspectives and methods for the security assessment of random number generators.