lists.openwall.net   lists  /  announce  owl-users  owl-dev  john-users  john-dev  passwdqc-users  yescrypt  popa3d-users  /  oss-security  kernel-hardening  musl  sabotage  tlsify  passwords  /  crypt-dev  xvendor  /  Bugtraq  Full-Disclosure  linux-kernel  linux-netdev  linux-ext4  linux-hardening  PHC 
Open Source and information security mailing list archives
 
Hash Suite for Android: free password hash cracker in your pocket
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Thu, 9 Jan 2014 16:21:52 -0200 (BRDT)
From: "Marcos Simplicio" <mjunior@...c.usp.br>
To: discussions@...sword-hashing.net
Subject: Re: [PHC] Lyra,
      Password Key Derivation Based On The Sponge Construction

Hi, everyone.

Some thoughts below. I'm a bit late answering (or maybe you were too fast
doing so :-) ), so I'm answering to the original post hoping to add some
ideas.


Both points are indeed relevant if the attacker has access to the KDF's
partial memory and access patterns, and I do believe Catena does a
wonderful job protecting against such attacks. Seriously, I do not think I
could do better. The only protection I see in Lyra is the fact that the
memory changes along the way, so the attacker needs to get the initial
values for getting any useful advantage (but, since there are potentially
many unchanged values if the memory usage is high, this is not a very
strong argument).

However, from a "risk analysis" point of view, I still think that the main
threat to KDFs is the usage of many processing cores and personalized
hardware performing brute force attacks. The reasoning is:

A) As far as I can tell, the attacker needs to be inside your system
*while you input the password* to access the internal memory used by the
KDF. Otherwise, how will he/she be able to see the reading patterns or the
memory initialized with the real password? So (1) this will not be a
majority of attackers and (2) what exactly is preventing him/her from
getting the plain password from the very start, since it will inevitably
be placed in memory at some point or be readable from an I/O device?
Protecting against this threat is definitely useful, but I believe that
treating this has lower priority than treating (B).

B) GPUs and FPGAs are a very serious threat already, especially because
brute-forcing a password does not require access to the system while the
password is being entered. All that is needed is the final KDF result,
obtained from a local storage (invaded at any time) or from an unprotected
communication channel. This is and will probably continue to be the most
likely attack venue against passwords. The only way I see to avoid such
attacks from being performed fast enough is to raise the associated costs
(number/power of GPUs and chip area of FPGAs)

There is an important difference with plain RC4 here, because RC4 normal
usage does provide the attacker with its internal state from the start as
its key stream, while KDFs never do so (the only thing the attacker gets
is the final result). Hence, KDFs are more similar to the (AFAIK, still
considered secure) RC4_Drop[n], in which the internal state is "worked on"
before it is provided to the outside world. Unless, of course, you are
considering threat (A) in your attacker model, in which case the "dropped"
data is available to the attacker.

Another thought is: if you build a KDF that always have the same (possibly
salt-dependent) visitation pattern, what prevents the attacker from
building a custom hardware that saves part of the memory states on a
cheaper memory devices (e.g., hard disks) and pre-fetches them when (and
only when) they are needed? I mean, the attacker knows what he will need
and when, so it should not be too hard to overcome the bandwidth
limitations of cheaper devices (something discussed both in Lyra and
scrypt's paper). I may be wrong, but it seems to me that protecting
against threat (A) by using password-independent visitation patterns
actually reduces the KDF's strength against threat (B)...

Well, those are my two cents :-)

BR,

Marcos.

>
> In message
> <CAOLP8p5wwnaOpPGW0rA+Q9nz-jYtKhEL0aujMALuRuG=8zQtRg@...l.gmail.com>
> , Bill Cox writes:
>
>>Thanks for this very interesting link.  Lyra first fills a matrix with
>> hash
>>data which is derived from the password, and then randomly picks a "row"
>>and for each location it updates the hash state from the location's
>> value,
>>and then XORs into the location the next output of the hashing engine.
>
> Two things worry me about the general approach Lyra takes.
>
> My first thought was that this sounds vulnerable to the same issue
> RC4 suffers from:   It takes more entropy to "randomize" it properly
> than is typically available for the purpose.
>
> Lack of entropy is a major issue in any password context, and therefore
> I think it is wise to pay attention to the:
>
> 	bits of entropy
> 	---------------
> 	bits of state
>
> ratio not getting too low.
>
> The second thought is that a large memory footprint, desirable for
> all the reasons the Lyra presentation mentions, vastly increases
> the ways and means to discover what is going on through covert
> channels.
>
> So as a general principle, I'm personally not going to be very
> impressed by claims on the general form:
>
> 	"The ${datestructure} can be sized however large you like"
>
> even if it comes with a mathematically proof of ${something}, unless
> it also comes with plan for how it gets initialized with only limited
> entropy being available, and analysis of to what extent the access
> patterns may reveal its state.
>
> Poul-Henning
>
> --
> Poul-Henning Kamp       | UNIX since Zilog Zeus 3.20
> phk@...eBSD.ORG         | TCP/IP since RFC 956
> FreeBSD committer       | BSD since 4.3-tahoe
> Never attribute to malice what can adequately be explained by
> incompetence.
>


Powered by blists - more mailing lists