A Search for Good Pseudo-random Number Generators : Survey and Empirical Studies

2018-11-03

Computer Science Review

AI Summary

Existing PRNG evaluations are fragmented and lack cross-paradigm comparability, hindering evidence-based selection for real-world applications. Method: We systematically benchmark over 100 PRNGs—including linear congruential, chaotic, and cryptographically derived variants—within a unified, reproducible framework. Our evaluation integrates statistical testing (NIST STS, TestU01), entropy analysis, runtime profiling, and cryptographic strength verification to assess statistical quality, computational efficiency, and security simultaneously. Contribution/Results: We identify underappreciated generators—e.g., xoshiro256++—that achieve both high speed and superior statistical robustness. Critically, we expose severe statistical flaws in several widely deployed PRNGs across major software libraries. This work delivers the first large-scale, cross-paradigm, reproducible empirical benchmark for PRNGs, providing practitioners with actionable guidance for design, standardization, and production deployment.

Problem

• Evaluating and ranking 30 PRNGs using empirical tests
• Classifying PRNGs into three distinct generator types
• Testing PRNGs with Diehard, TestU01, and NIST suites

Innovation

• Empirical testing with Diehard, TestU01, NIST
• Classification into three PRNG architecture groups
• Performance ranking of 30 selected generators