@Article{jcs.2025.063529,
AUTHOR = {Issah Zabsonre Alhassan, Gaddafi Abdul-Salaam, Michael Asante, Yaw Marfo Missah, Alimatu Sadia Shirazu},
TITLE = {An Overview and Comparative Study of Traditional, Chaos-Based and Machine Learning Approaches in Pseudorandom Number Generation},
JOURNAL = {Journal of Cyber Security},
VOLUME = {7},
YEAR = {2025},
NUMBER = {1},
PAGES = {165--196},
URL = {http://www.techscience.com/JCS/v7n1/62944},
ISSN = {2579-0064},
ABSTRACT = {Pseudorandom number generators (PRNGs) are foundational to modern cryptography, yet existing approaches face critical trade-offs between cryptographic security, computational efficiency, and adaptability to emerging threats. Traditional PRNGs (e.g., Mersenne Twister, LCG) remain widely used in low-security applications despite vulnerabilities to predictability attacks, while machine learning (ML)-driven and chaos-based alternatives struggle to balance statistical robustness with practical deployability. This study systematically evaluates traditional, chaos-based, and ML-driven PRNGs to identify design principles for next-generation systems capable of meeting the demands of high-security environment like blockchain and IoT. Using a framework that quantifies cryptographic robustness (via NIST SP 800-22 compliance), computational efficiency (throughput in numbers/sec), and resilience to adversarial attacks, we compare various PRNG architectures. Results show that ML-based models (e.g., GANs, RL agents) achieve near-perfect statistical randomness (99% NIST compliance) but suffer from high computational costs (10³ numbers/sec) and vulnerabilities to model inversion. Chaos-based systems (e.g., Lorenz attractor) offer moderate security (97%–98% NIST scores) with faster generation rates (1.5–2 ms/number), while traditional PRNGs prioritize speed (10<sup>6</sup> numbers/sec) at the expense of cryptographic strength (70%–95% NIST pass rates). To address these limitations, we propose hybrid architectures integrating classical algorithms with lightweight ML components (e.g., TinyGANs), reducing training time by 40% while maintaining 98% NIST compliance. This work provides three key contributions. These include a quantitative security-speed trade-off analysis across PRNG paradigms, cryptanalytic insights into ML-PRNG vulnerabilities (e.g., overfitting to training data), and actionable guidelines for optimizing hybrid designs in resource-constrained settings. By bridging deterministic and stochastic methodologies, this study advances a roadmap for developing adaptable, attack-resistant PRNGs tailored for evolving cryptographic standards.},
DOI = {10.32604/jcs.2025.063529}
}