Learned Pseudo-Random Number Generator Based on Generative Adversarial Networks

Pseudorandom number generators (PRNGs) are fundamental components of modern cryptography and information security. Due to the inherent complexity and unpredictability of neural networks, they have become an attractive alternative for designing PRNGs. In recent years, several PRNGs based on Generative Adversarial Networks (GANs) have been proposed, which is an end-to-end generation approach. However, current GAN-based PRNGs have been found to suffer from several limitations. For instance, their input length is often too short to resist exhaustive attacks. And the randomness tests of NIST are not always met under recommended parameter settings, indicating the need for further improvement in the design of GAN-based PRNGs. To address these issues, this paper proposes a practical and secure PRNG designed based on GANs. Specifically, we rationalized the input and output parameters by setting the input to 32 8-bit unsigned integers and the output to 256 8-bit unsigned integers in a single iteration, ensuring that the input data space reaches \(2^{256}\) . We also optimized the GAN network architecture and incorporated the GELU activation function to ensure that the generated output passes all randomness tests of NIST under the recommended parameter settings.