> "Silicon has no secrets."

It would be very interesting to examine some of the DES Cracker gate
array chips with these tools.  Though the chips worked great in
simulation, and each search engine came from exactly the same VHDL
source code, some number of the 24 search engines on each manufactured
chip have subtle errors that corrupt some part of the DES calculation,
causing <1% of both false-negatives and false-positives when doing
brute force key search.  Particular patterns would rapidly fail.  Each
chip also failed its standard test vectors after manufacturing.  We
accepted and used the chips anyway, working around the issues by
writing fault-tolerant software, because we couldn't afford to do
another 6- to 8-week chip fabrication turnaround (plus the time needed
to characterize and fix the problem).

Was this subtle error an introduced malicious change?  A bug in the
chip compiler?  Or was it caused by a poorly characterized process
variation in the factory?  We never found out.

A flakey DES cracker that worked on some keys and not on others would
have made less of a public policy impact.  We might never have
announced it at all, if it was failing in ways we couldn't understand
and couldn't compensate for.  Based on unattributable stories we heard
from other cryptographers, we had been concerned from that if we did
the work publicly, NSA would figure out a way to screw us (by leaning
on the chip vendor through some executive-level contact: "Do this for
the good of your country", then they'd refuse to make our chips, or
break the chips somehow).

Luckily, since we had a mix of working and non-working engines, we
could adjust the software to use all engines in the first pass, but
then only use known-good search engines to re-search blocks of keys
that had been searched by known-flakey search engines.  The result was
a cracker that works at its designed speed for most keys and
plain/ciphertext combinations, but which works about 40% slower on <1%
of such combinations.

I suspect error rather than malice here, but this raises an
interesting issue.  Even if you find a chip design bug that lets
people break in, can it be designed to LOOK LIKE error rather than
malice?  Clearly the data bus comparator and reserved cache line thing
from LEET would show up as malice, if ever discovered.  But there are
subtler and perhaps harder to exploit changes that would leave much
lighter traces of deliberate sabotage.


PS:  Lots of security bugs in software look like error, yet we know
that some are malicious (like the one briefly introduced into the Linux
kernel sources by breaking in to one of the machines holding the source
tree):  http://kerneltrap.org/node/1584

The Cryptography Mailing List
Unsubscribe by sending "unsubscribe cryptography" to [EMAIL PROTECTED]

Reply via email to