Assessing Cryptographic Random Number Generators using Machine Learning

Abstract

This thesis investigates the intersection of machine learning and cryptography, with a particular emphasis on the capacity of neural networks to model cryptographic hash functions and evaluation of randomness in binary sequences. The feasibility of training neural networks to replicate the behavior of cryptographic hash functions, focusing on a reduced variant of the SHA-3 algorithm. Empirical results indicate that while neural networks can accurately approximate steps of the Keccak- permutation, they exhibit limited generalization capability across multiple rounds of the Keccak- function. In the second phase, the focus shifts to the application of machine learning techniques for the analysis of randomness in binary sequences. Utilizing transformer-based architectures, thesis demonstrates that these models can achieve high predictive accuracy on the final bit of a sequence, including those classified random by conventional statistical test suites such as NIST SP800-22. These findings suggest that machine learning models may serve as practical complementary tools to traditional statistical methods, offering a novel approach for uncovering subtle, exploitable patterns that dodge standard randomness assessments.This thesis investigates the intersection of machine learning and cryptography, with a particular emphasis on the capacity of neural networks to model cryptographic hash functions and evaluation of randomness in binary sequences. The feasibility of training neural networks to replicate the behavior of cryptographic hash functions, focusing on a reduced variant of the SHA-3 algorithm. Empirical results indicate that while neural networks can accurately approximate steps of the Keccak- permutation, they exhibit limited generalization capability across multiple rounds of the Keccak- function. In the second phase, the focus shifts to the application of machine learning techniques for the analysis of randomness in binary sequences. Utilizing transformer-based architectures, thesis demonstrates that these models can achieve high predictive accuracy on the final bit of a sequence, including those classified random by conventional statistical test suites such as NIST SP800-22. These findings suggest that machine learning models may serve as practical complementary tools to traditional statistical methods, offering a novel approach for uncovering subtle, exploitable patterns that dodge standard randomness assessments.