view libtomcrypt/notes/tech0002.txt @ 629:bdadc7f4b97d

Set IPTOS_LOWDELAY for IPv6 too
author Matt Johnston <matt@ucc.asn.au>
date Wed, 26 Oct 2011 16:02:06 +0000
parents 1b9e69c058d2
children
line wrap: on
line source

Tech Note 0002
How to avoid non-intrusive timing attacks with online computations
Tom St Denis

Introduction
------------

A timing attack is when an attacker can observe a side channel of the device (in this case time).  In this tech note
we consider only non-intrusive timing attacks with respect to online computations.  That is an attacker can
determine when a computation (such as a public key encryption) begins and ends but cannot observe the device 
directly.  This is specifically important for applications which transmit data via a public network.

Consider a Diffie-Hellman encryption which requires the sender to make up a public key "y = g^x mod p".  Libtomcrypt
uses the MPI bignum library to perform the operation.  The time it takes to compute y is controlled by the number
of 1 bits in the exponent 'x'.  To a large extent there will be the same number of squaring operations.  "1" bits in
the exponent require the sender to perform a multiplication.  This means to a certain extent an attacker can 
determine not only the magnitude of 'x' but the number of one bits.  With this information the attacker cannot directly
learn the key used.  However, good cryptography mandates the close scrutiny of any practical side channel.

Similar logic applies to the other various routines.  Fortunately for this case there is a simple solution.  First, 
determine the maximum time the particular operation can require.  For instance, on an Athlon 1.53Ghz XP processor a
DH-768 encryption requires roughly 50 milliseconds.  Take that time and round it up.  Now place a delay after the call.

For example, 

void demo(void) {
   clock_t t1;

   // get initial clock
   t1 = clock();
   
   // some PK function
   
   // now delay 
   while (clock() < (t1 + 100));
   
   // transmit data...
   
}

This code has the effect of taking at least 100 ms always.  In effect someone analyzing the traffic will see that the
operations always take a fixed amount of time.  Since no two platforms are the same this type of fix has not been 
incorporated into libtomcrypt (nor is it desired for many platforms).  This requires on the developers part to profile
the code to determine the delays required.

Note that this "quick" fix has no effect against an intrusive attacker.  For example, power consumption will drop
significantly in the loop after the operation.  However, this type of fix is more important to secure the user of the 
application/device.  For example, a user placing an order online won't try to cheat themselves by cracking open their
device and performing side-channel cryptanalysis.  An attacker over a network might try to use the timing information
against the user.