Re: advice sought on key/data histogram analysis of rijndael/128 and serpent
From: lkcl (lkcl_at_lkcl.net)
Date: 18 Oct 2005 04:08:45 -0700
in my intial tests, i was running with number of blocks
equal to several thousand - sufficient to produce
1 million bits.
i found that even then, there were an inordinately
high number of p-values > 0.01.
this was what alerted me to the problem in the first
actually, it was NIST's STS bombing out with
a coredump because the balance of 1s and 0s was
_so_ way outside the statistical 50% balance mark
that there was a buffer overrun (inside the "runs"
test iirc - this _was_ two years ago, with version 15).
i continued to reduce the block size, and have found
that it's even possible to have only one block, and for
there to be an inordinate number of failures - i mean
we're talking 3.5% here! an average of 70,000
p-values < 0.01 out of 2048000 tests.
hmmm... *thinks*.... okay, that's coincidentally close
to the value "4.5/128". .... hmmm...
okay, what's 4.5/128 * 2048000 that's 72000.
hmmm... okay, what's a better appx median in one
of the histogram outputs: it's 71300. hmmm....
perhaps i should try increasing the number of blocks
being tested, again...