Generative Adversarial Networks-Based Pseudo-Random Number Generator for Embedded Processors

摘要

A pseudo-random number generator (PRNG) is a fundamental building block for modern cryptographic solutions. In this paper, we present a novel PRNG based on generative adversarial networks (GAN). A recurrent neural network (RNN) layer is used to overcome the problems of predictability and reproducibility for long random sequences, which is found in the result of the NIST test suite for the previous method. The proposed design generates a random number of 1,099,200-bits with a 64-bit seed. The proposed method is also efficiently implemented on embedded processors by using the Edge TPU. To support the Edge TPU, the proposed GAN based PRNG is converted to a Tensor-Flow Lite model. During model training, the number of epochs is significantly reduced with the proposed approach. The PRNG generates random numbers in 13.27ms using the Edge TPU. Also, our PRNG achieved a speed of 1.0GB/s, which is about 6.25x compared to the speed of other lightweight PRNG. To the best of our knowledge, this is the first GAN based PRNG for embedded processors. Finally, generated random numbers were tested through the NIST random number test suite. Compared with the previous method, the proposed method reduced the percentage of test failures by 2.85x. The result shows that the proposed GAN-based PRNG achieved high randomness even on embedded processors.