Testing Cryptographically Secure Pseudo Random Number Generators with Artificial Neural Networks

We present a new way of testing Random Number Generators (RNGs). Our approach allows to test Pseudo Random Number Generators (PRNGs) including Cryptographically Secure Pseudo Random Number Generators (CSPRNGs). The paper describes how to use machine learning for this. To construct a tester we compare the properties of three most common learning techniques to find the one most suitable one for testing RNGs. By analyzing the system during training and regarding the expected behavior of random numbers, we define an optimizer for learning RNGs. Based on the results and regarding the behavior of the machine learning algorithm, we define a rating for RNGs. On a state-of-the-art GPU cluster, we evaluate the full tester for multiple PRNGs. Additionally, we compare the results with the results from the commonly used test suite dieharder. The results prove that the developed tester is suitable for testing random numbers. In comparison to dieharder, it is even more powerful and able to replace it. Our tester could disclose weaknesses in PRNGs that are wrongly considered as CSPRNG. This could increase the security of many cryptographic protocols based on random numbers.