3
$\begingroup$

I am using the NIST suite to test the randomness of binary numbers and I have a 64KB dataset. I am confused with the results in finalAnalysisReport.txt.

I have the following questions:

  1. does the test pass only if there is no "*" in both "p-value" and "proportion" columns? example:
12 5 7 3 5 4 5 3 4 2 0.108791 45/50 * NonOverlappingTemplate
  1. Does the "p-value" only need to be greater than 0.0001 to pass? enter image description here
  2. If I can't set the same parameters so that all tests pass, is it allowed to run 15 separate tests?
Improve this question
$\endgroup$
3
  • 2
    $\begingroup$ Why such a small dataset? NIST runs better on larger datasets! $\endgroup$
    – Maarten Bodewes ♦
    Commented Dec 28, 2023 at 22:38
  • $\begingroup$ @ Maarten Bodewes♦ Thank you! I understand, but I'm limited by the size of the SRAM, do you have any other way to expand the true random number? Like writing a code or something? $\endgroup$
    – yuting zhang
    Commented Dec 29, 2023 at 0:12
  • 1
    $\begingroup$ If you can export it you may be able to stream it, put it into EEPROM etc.. You could also call the RNG multiple times, but I don't know if that will influence the state of the RNG; it depends on what you are trying to test in that case. I'm not familiar with your runtime environment. $\endgroup$
    – Maarten Bodewes ♦
    Commented Dec 29, 2023 at 0:16

1 Answer 1

Reset to default
3
$\begingroup$

1.) The pass/fail threshold is printed at the bottom of the finalAnalysisReport.txt file. It looks like:-

The minimum pass rate for each statistical test with the exception of the random excursion (variant) test is approximately = x for a sample size = y binary sequences.

45/50 is a fail.

2.) No. $\alpha = 0.01$ is for the individual NIST tests. That seems reasonable given the relatively huge sample sizes. But the final $\chi^2$ test seems completely unreasonable with $\alpha = 0.0001$. It's almost as if NIST are encouraging poor generators to pass whereas common statistical analyses would use an $\alpha = 0.05$. Cui bono?

But. I am somewhat doubtful of the NIST tests. NIST/NSA objectives are not your objectives. And NIST are not great coders. I've just gotten:-

10   0   0   0   0   0   0   0   0   0  0.000000 *    0/10   *  Universal

for a 1 MB test from /dev/urandom. Zero point zero?

3ish.) 64 kB is much too low for a reliable NIST SPS test. So I suggest you generate more entropy or use other test suites more appropriate for smaller samples. There is ent and it's contemporary replacement ent3000.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.