Date: Fri, 10 Jul 2015 06:54:10 -0700
From: Bill Cox <waywardgeek@...il.com>
To: "discussions@...sword-hashing.net" <discussions@...sword-hashing.net>
Cc: John Tromp <john.tromp@...il.com>
Subject: Data port hardness in key stretching and PoW?

I think Alexander used the term "port-hardness" a couple of times, but I
did not see this topic analyzed anywhere.  How secure is a scheme for
password hashing or PoW which depends on the high cost of transmitting and
receiving data over IC I/O ports?

I've been trying to figure out why my design for a massively parallel ASIC
+ router attack on the Momentum PoW system will not easily reduce system
cost.  I can increase the workers arbitrarily without adding significant
amounts of RAM, so the memory*time cost can be arbitrarily reduced.  This
proves that Momentum is not sequential-memory-hard, or as we generally say
on this list, just "memory-hard".  However, when I tried to figure out how
to build the system cheaply, I ran into the high cost of routing chips with
high speed interconnect.  DRAM, which not cheap, is still comparatively
cheaper than routing chips.

To execute Momentum fast, my single-CPU algorithm writes hashes to external
memory once, and reads it once.  With a decent on-chip hardware hashing
core, which costs essentially nothing, this system will be memory bandwidth
bound rather than CPU compute time bound or external memory latency bound.
All of my favorite memory-hard algorithms in  this competition are
similarly memory bandwidth bound, at least in some modes of operation (this
is true for Yescrypt, Lyra2, and Argon2, and with the right choice of hash
function, also Catena).

I was not able to prove that Momentum requires any particular threshold of
data to be transmitted across IC boundaries, so I can't yet prove it's
"port-hard".  However I can easily prove this for an EARWORM-like password
hashing scheme.  In particular, we can use a large ROM and small RAM for a
PoW system with reasonably fast verification, as I posted recently.  There
is no way to generate the password hashes without bringing together the RAM
data and ROM data, requiring a minimum amount of data to be transmitted.
Any parallel attack on such a system will have to transmit the ROM data
and/or RAM data over I/O ports to complete hashing.  While the expense of
the large ROM can be shared among workers, the I/O bandwidth to the ROM
increases linearly with the number of workers.  The algorithm is not
memory-hard, but it is port-hard.

What makes a port-hard algorithm hard to attack in a custom ASIC?  I know
that most ASICs have their I/Os around the edge, and thus the chip area is
proportional to the square of the number of pads.  It is also difficult to
design high-speed serial ports, and the IC's available with such ports are
pricey.  However, IBM has a technology for distributing pads uniformly
throughout the die, though there is a large unusable area of silicon around
each, IIRC.  Also, as the number of pins on a chip grows, the package cost
can get out of control.  High-end flip-chip packaging required for IBM's
technology never did come down enough in cost to make them attractive for
anything but very high-end systems, as far as I know.

Can we build a secure password hashing scheme or PoW based on port-hardness?

Bill

Thanks,
Bill

Content of type "text/html" skipped

Powered by blists - more mailing lists