• PDF / 1,839,982 Bytes
• 19 Pages / 439.37 x 666.142 pts Page_size
• 48 Downloads / 242 Views

DOWNLOAD

REPORT

---

Pseudo‑random number generation using LSTMs Young‑Seob Jeong1 · Kyo‑Joong Oh2 · Chung‑Ki Cho1 · Ho‑Jin Choi2

© Springer Science+Business Media, LLC, part of Springer Nature 2020

Abstract Previous studies have developed pseudo-random number generators, where a pseudo-random number is not perfectly random but is practically useful. In this paper, we propose a new system for pseudo-random number generation. Recurrent neural networks with long short-term memory units are used to mimic the appearance of a given sequence of irrational number (e.g., pi), and these are intended to generate pseudo-random numbers in an iterative manner. We design algorithms to ensure that the output sequence contains no repetition or pattern. Through experimental results, we can observe the potential of the proposed system in terms of its randomness and stability. As this system can be used for parameter approximation in machine learning techniques, we believe that it will contribute to various industrial fields such as traffic management and frameworks for sensor networks. Keywords  Pseudo-random number generation · Recurrent neural networks · SHA2 · Irrational number · NIST test suite

1 Introduction The fourth industrial revolution has been augmented with significant improvements in machine learning techniques and an exponentially increasing amount of available data. Machine learning techniques, especially deep learning models, play a crucial role in elevating artificial intelligence above a human level in * Ho‑Jin Choi [email protected] Young‑Seob Jeong [email protected] Kyo‑Joong Oh [email protected] Chung‑Ki Cho [email protected] 1

Bigdata Engineering Department, Soonchunhyang University, Asan, ChungNam, South Korea

2

School of Computing, KAIST, 291 Daehak‑ro, Yuseong‑gu, Daejeon, South Korea



13

Vol.:(0123456789)



Y.-S. Jeong et al.

certain areas (e.g., the game of Go or general knowledge quizzes). They are also employed in various industrial fields, such as traffic identification and frameworks for sensor networks  [15, 35]. Such techniques are usually trained via stochastic processes, which are practically related to random number generation. A stochastic process is associated with a collection of random variables, and these random variables are sampled for particular iterations. The value of each random variable is usually sampled based on a distribution over the possible values, and this is where random number generation is utilized. For instance, given a particular n-dimensional cumulative multinomial distribution, the sampling process is performed by generating a random number between zero and the nth value of the cumulative distribution. Random number generation is also important in the field of security and encryption [25, 26], where it is used to generate key values for the data encryption standard algorithm (DES)  [19] or advanced encryption standard algorithm (AES) [20]. It is obvious that the quality of the random number generation module will strongly influence the quality of a security service, training of