A Machine Learning-Based Online Evaluation Method for Statistical Properties of Random Number Generators

Random number generators are influenced by internal and external factors, leading to reduced randomness during work; therefore, real-time quality detection of these generators is crucial. However, existing international statistical suites detect statistical properties of random numbers offline, and thus cannot reflect real-time randomness. This paper proposes an on-chip detection method for the statistical properties of random numbers using machine learning. By collecting raw random numbers from ring oscillators, a type of true random number generator, we established a real-time online quantification evaluation mechanism for entropy source quality based on the correlation between the prediction accuracy of the machine learning methods and the minimum entropy. The results show a significant negative correlation between minimum entropy and prediction accuracy for the ring oscillator entropy sources with different ring counts, with a Pearson correlation coefficient of approximately −0.95 and a P-value of 6.98e-07. Furthermore, the Long Short-Term Memory (LSTM) network achieves higher prediction accuracy compared to feedforward Neural Network (FNN), Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU), and Broad Learning Systems (BLS), with a parameter size of 17.04 KB. Thus, it is feasible to measure entropy source quality with the network prediction accuracy. Moreover, the small network size makes it suitable for on-chip deployment, offering promising application prospects.