On 11/20/2011 08:10 AM, Piotr Szturmaj wrote:
bcs wrote:
On 11/04/2011 04:27 AM, Piotr Szturmaj wrote:
bcs wrote:
Are you re-implementing the function kernels your self or are you using
an existing implementation? Given what I've heard about things like
side-channel attacks using processing times to recover keys, I'd rather
not see Phobos use anything written by less than the best expert
available.
Until now, I implemented some hash functions. There are no branching
instructions in their transform() routines, so theoretically processing
time is independent of the function input.
From my very incomplete memory I found the source I was looking for (I
googled for "aes interperative dance") here
http://www.moserware.com/2009/09/stick-figure-guide-to-advanced.html
Look for "Foot-Shooting Prevention Agreement" in one of the images
~20-25% of the way down.
tl;dr; It mentions "cache-based, timing, and other side channel
attacks". Unless you can explain to me what those are, in painful
detail, I don't think we should trust you to avoid them. Get a good
vetted C implementation and wrap it with a nice D API and call it a day.
>
Sorry for late answer.
Np, but I still have a number of concerns:
What is the advantage to implementing the kernels of any of these
functions in D? Will the code be faster? Smaller? More secure? More
maintainable? (On the other hand, the value of doing the API code in D
goes with no debate.)
How many people in the D community have the experience and know-how to
review the security of an implementation? If there are less than 2 or 3
people who can do that, can we afford to include native kernels? We
can't have just one and if we have only two and one leaves for some
reason the code becomes un-maintainable for lack of a reviewer. *I*
wouldn't be comfortable with less than about 4-5.
I didn't implement AES yet, but it can be implemented without lookup
tables (countermeasure to cache-timing attacks).
Branch-free code without secret-data dependent memory accesses is free
from cache-timing vurnelabilities on most architectures. The remaining
are architectures where instruction execution times (cycles) may depend
on data (f.i. multiply by 0 may take 1 cycle, and >2 cycles when
multiplying by other numbers). I don't know any concrete examples but I
realize such cycle-varying instructions _may_ exist. Of course such
instructions must be avoided in secure code.
The most hard to avoid side channel attack is Power Analysis. On almost
all CPUs instruction power drain depends on the operands. Again
multiplying by 0 consume less power than multiplying by other numbers
and that may be distinguished on the oscilloscope. There are some
countermeasures including operand blinding, but not all crypto
algorithms support that.
I only implemented some hashing functions: MD5, SHA1/224/256/384/512 and
Tiger1/2. MD5 and SHA have code that is branch-free and does not use any
data dependent lookups so their execution time should be constant,
making timing attacks harmless.
Unfortunately Tiger function does lookups based on the source bytes and
is vurnelable to cache-timing attacks. This may be considered insecure
and removed (openssl does not support it either) as making it secure may
be not worth it.
The only problem that remain is the power analysis which require
physical access to the computer. We need to ask ourselfs to what degree
we must secure our code. I'm going to argue that no single
implementation is secure to power analysis on all architectures. I think
we must stick to cache/branch timing level and forget about power
analysis as its scope is beyond D's specification. We simply can't avoid
most power analysis attacks because most of the countermeasures belong
to the hardware instead of the software level.
