Message-ID: <CAOLP8p5EYwvQZWiXd7EO5AGrGBzpHz+2O-_5xGZDvFcV=Wc=2Q@mail.gmail.com>
Date: Wed, 26 Feb 2014 09:35:17 -0500
From: Bill Cox <waywardgeek@...il.com>
To: discussions@...sword-hashing.net
Subject: Re: [PHC] Should we care about "parameter influence" attacks against PBKDF2?

On Wed, Feb 26, 2014 at 4:14 AM, Solar Designer <solar@...nwall.com> wrote:
> As you probably realize, hashing of all the input parameters is not the
> only way to prevent this attack.  A final step may (or may not,
> depending on how it's chosen) also prevent the attack.  (Arguably, this
> translates into hashing of the iteration count anyway, via the iteration
> count indirectly affecting the input to this final crypto hash step.)

This is what I was thinking.  I currently call PBKDF2 at the end, but
the only parameter I pass in is the hash.  If input parameters cause
correlations of any sort in the hash just before this step, the final
hash should eliminate them.  With this, there isn't any reason to hash
the input parameters, is there?  As you say, we could argue that the
final hash *is* the hash of the input parameters.  I like this
solution better than hashing a "tweak" value in at the beginning.

I'm digging a bit more into the design of PBKDF2 and HMAC now.  HMAC
is cool, and seems like the right way to hash the initial password,
salt, and data.  PBKDF2, with c == 1 (iteration count), and output
length == 32 turns into just a call to HMAC, followed by a seemingly
useless additional call to HMAC with the hash and a 1-byte 0 as input.
 I am considering calling HMAC directly instead of PBKDF2 in the
beginning, though the extra HMAC call is no big deal.  Is it worth
calling PBKDF2 simply to give people peace of mind without having to
convince them that it's OK to just call HMAC?

I also currently use PBKDF2 for the final hash, but I'm thinking of
abandoning that and using just the crypto-hash directly.  If a user
asks for a 64-byte output, provides a 64-byte password and a 64-byte
salt, and if I used a 64-byte hash like Blake2b throughout, then a
user is getting a bit of raw deal from PBKDF2.  He is probably
expecting a full 512 bits worth of security, but PBKDF2 only gives him
256, even though it takes 512 bit inputs and delivers a 512 bit
result.  This is because it compresses the hash data down to 256 bits,
and then uses that state to generate the output for any length.

I can't figure out why a 512 bit result hash is any more secure than a
256 bit hash for any realistic application, but if the user asks for
it, I feel bad giving him the output from PBKDF2.  At the same time, I
don't want to take on the challenge of using a small width hash
function to securely hash the final wide width result.  It seems too
easy to mess up.  For example, just calling Blake2s on two 32-bit
hashes results in a 64-byte hash where changing 1 input bit only
effects 32 bytes of the output.  A wide hash needs to be better than
that.  I'm tempted to use a 512 bit hash instead, like Blake2b, and
simply not support anything wider.

> Would this fall under "the KDF designer messed up" (if such a final step
> is missing or does not prevent the attack)?

Yes, that's what I meant.  When Steve posted that devastating attack
on my NoelKDF hash, it scared the heck out of me.  Now there is no way
I'm going to deliver a KDF that does not call a secure hash function
one final time on the output.  Even PBKDF2 messed up in this case,
delivering the XOR of successive rounds without a final HMAC call on
the result.  It's really scary how hard this stuff is to get right.  A
final secure hash on the entire output seems like a good measure to
me.

As a side-effect, having a
> final step like this also prevents hash upgrades to higher iteration
> count, which brings us right to:

I don't think it prevents application of "garlic" the way Catena does
it.  Catena (and my KDF since I copied Caten), applies a final hash
inside the garlic loop, so it gets called garlic times, but that's no
biggie.

A concern I have is that in "server relief" mode, Catena (and my KDF,
again because I copied them), the final application of a secure hash
function is skipped, and the hash just before this step is returned to
the caller.  In the way this is meant to be used, that returned hash
is just as sensitive as the password, since it can be used by an
attacker instead of the password to login to a server.

Because of this, I was thinking that it is OK if this returned hash
has weaknesses that an attacker could use to more quickly guess the
password, but an attackers real goal might be to guess the password
hoping to use it to log into other sites.  Therefore, I've got work to
do to strengthen server relief mode :-)

Catena also returns the hash value in server relief mode before the
final hash is called, but in the master branch, all hashing is done
with a cryptographically secure hash function, so the previous hash
should be fine.  In Catena, the only reason for the final hash seems
to be to enable server-relief mode in a garlic-compatible manner.
However, in my "waywardgeek" branch of Catena, I introduced a
"FastHash" option for fast but cryptographically weak memory hashing
(which is still the right thing to do, IMO).  I guess I've broken his
server relief mode...

I don't know if it's useful for me to fix the waywardgeek branch,
since I'm not seeing significant activity in the Catena git
repository.  If it were to move forward, I'd upgrade the fast hash
functions to be more SIMD friendly and base it on ARX (addition,
rotation, XOR), drop the multiplications, and probably write SSE and
AVX2 versions.  There's probably another factor of 2X or more in speed
to be gained vs my last benchmarks on my waywardgeek branch.  I'd also
need to fix this server relief problem I've introduced.  However,
there's only a few weeks left to go, and the master branch still is
10X too slow to be competitive, when it's sooo easy to upgrade...

Thanks for the description of these issues.  It makes a lot more sense now.

Bill

Powered by blists - more mailing lists