Informatics Colloquium 15.10. Revisiting statistical tests for random data analysis
Informatics Colloquium 15.10. 2019, 14:00 lecture hall D1
Mgr. Marek Sýs, Ph.D., FI MU
doc. RNDr. Petr Švenda, Ph.D., FI MU
prof. RNDr. Václav Matyáš, M.Sc., Ph.D., FI MU
Revisiting statistical tests for random data analysis
Abstract: Random data play important role in many areas e.g. gaming, simulation,
cryptography, etc. The quality of the data is critical for several areas hence
analysis of used generators (true or pseudo random) is crucial there. Multiple
incompatible test suites exist, with the tests that analyse different but
similar patterns. We created a framework to comfortably run all tests from the
commonly used statistical testing batteries (NIST STS, Dieharder, TestU01 and
BoolTest). Using this tool, we analyzed the output of 66 cryptographic functions
and evaluated the power/uniqueness of individual tests in the domain of
cryptographic data. Surprisingly, a subset of only 14 tests would be enough to
detect all the biases detected by the whole set -- yet no standard battery
contained them all. In the talk, we will discuss the results we obtained in more
details. Also, we will introduce our BoolTest battery that outperforms standard
batteries when partial bits on fixed positions in data blocks are correlated.