This is a humble try at determining whether a piece of code runs in constant time or not. The approach is easy: run it with different inputs, measure execution time and apply statistics. This tool is fairly small: the relevant code is around 350 lines.
The approach is described in our paper:
Oscar Reparaz, Josep Balasch and Ingrid Verbauwhede
dude, is my code constant time?
A C compiler.
To build, run
make. This builds a bunch of binaries named
Let's have a look at
memcmp()-based comparison of passwords or MAC tags.
It fails very quickly:
$ ./dudect_cmpmemcmp_-O2 meas: 0.37 M, max t: +1271.13, max tau: 3.47e-03, (5/tau)^2: 2.07e+06. Definitely not constant time.
The output says: the tested function was executed 370k times ("measurements") and we got a t statistic value of 1271, deeming the implementation not constant time. t values larger than 5 mean there is very likely a timing leakage. (The other figures are not so relevant now.)
Now try some crypto. This is a 32-bit AES128 implementation:
$ ./dudect_aes32_-O2 meas: 0.59 M, max t: +561.16, max tau: 9.58e-04, (5/tau)^2: 2.72e+07. Definitely not constant time.
In this case, the output says that after 590k measurements we got a t statistic of 561. A value of 561 is overwhelming evidence that there is timing leakage.
The two previous cases were easy to catch, since the timing leaks are huge. There are some cases that are a bit more elaborated. In those cases, the tool may not detect at first the timing leak, but only when enough measurements are accumulated. For example let's try a curve25519 implementation with an intentional timing leak:
$ ./dudect_donnabad_-O2 meas: 0.00 M, not enough measurements (9000 still to go). [...] meas: 0.01 M, max t: +0.36, max tau: 3.59e-05, (5/tau)^2: 1.94e+10. For the moment, maybe constant time. meas: 0.02 M, max t: +8.22, max tau: 4.02e-04, (5/tau)^2: 1.55e+08. For the moment, maybe constant time. [...] meas: 0.02 M, max t: +11.35, max tau: 5.63e-04, (5/tau)^2: 7.89e+07. Probably not constant time. meas: 0.02 M, max t: +16.38, max tau: 7.91e-04, (5/tau)^2: 4.00e+07. Probably not constant time. meas: 0.02 M, max t: +23.71, max tau: 1.20e-03, (5/tau)^2: 1.73e+07. Probably not constant time. ^C
(I Ctrl-C'ed after some minutes.) Here is what happened:
If the code exhibits timing variability, the t statistic grows as more measurements are available. If it surpasses 10 then most likely there is some timing leak.
If there is no leak, the t statistic will not go beyond 10, for whatever number of measurements. This is the case when testing a constant-time piece of code:
$ ./dudect_cmpct_-O2 meas: 1.00 M, max t: +1.94, max tau: 1.94e-06, (5/tau)^2: 6.64e+12. For the moment, maybe constant time. meas: 2.00 M, max t: +1.85, max tau: 9.27e-07, (5/tau)^2: 2.91e+13. For the moment, maybe constant time. meas: 3.00 M, max t: +1.56, max tau: 5.20e-07, (5/tau)^2: 9.24e+13. For the moment, maybe constant time. meas: 4.00 M, max t: +1.63, max tau: 4.08e-07, (5/tau)^2: 1.50e+14. For the moment, maybe constant time. meas: 4.98 M, max t: +1.22, max tau: 2.45e-07, (5/tau)^2: 4.15e+14. For the moment, maybe constant time. meas: 5.98 M, max t: +1.32, max tau: 2.20e-07, (5/tau)^2: 5.14e+14. For the moment, maybe constant time. meas: 7.00 M, max t: +1.52, max tau: 2.17e-07, (5/tau)^2: 5.33e+14. For the moment, maybe constant time. meas: 7.96 M, max t: +1.70, max tau: 2.13e-07, (5/tau)^2: 5.52e+14. For the moment, maybe constant time. meas: 9.00 M, max t: +1.19, max tau: 1.33e-07, (5/tau)^2: 1.42e+15. For the moment, maybe constant time. [...] meas: 70.71 M, max t: +2.78, max tau: 3.93e-08, (5/tau)^2: 1.62e+16. For the moment, maybe constant time. ^C
Here the t statistic will never go beyond 10, since the code is constant time. The tool runs until it sees enough evidence that the code is not constant time. This means that if the code is constant time, it will run forever. Just Ctrl-C it when you think you are done.
dudect_aes32 T-tables implementation of the AES.
A nice example of variable-time crypto implementation.
constant-time AES implementation.
dudect_donna Langley's curve25519-donna. Intended to yield
dudect_donnabad Variant with a glaring timing leak.
dudect is distributed as a single-file library for easy building.
dudect.hto your include directories
#include "dudect.h"from your source files.
dudect_main()from your main function
Whether some piece of code executes in constant time depends on lots of
factors, such as: how the code is written, compiler version,
compiler flags, architecture, microcode version, phase of the moon,
etc. To see how different compiler optimization levels affect this,
compare the artifact
dudect_simple_O0 (no optimization, runs in variable
time in a 2019 MacBook) vs
dudect_simple_O2 (optimized, can't
detect leakage with a few million measurements on same platform).
How does this work? In a nutshell, this code takes two different inputs, runs the piece of code many times for each input and measures how much time it takes. If there is a statistical difference on the (average) time it takes to run with different inputs, the implementation is deemed not time constant. For details, read our paper or have a look at the source.
Is this a timing attack? No. This is leakage detection. Presence of leakage is necessary, but not sufficient for an attack to work.
My code passes this tests. Does it mean it is really time constant? Absolutely not. The test implemented is perhaps the simplest form of leakage detection. For serious assessment, you should also run many other tests, with specially crafted input vectors. The test harness implemented here is not yet that comprehensive. You're encourage to submit an implementation that is not constant time yet passes the test--in that way we can improve the test suite. The more you know about the implementation, the better you can design input vectors. For inspiration, see these RSA test vectors.
This was done before. Sure. For example, see the previous work of Barthe et al. or Langley. Here we take another, very different approach. We do not try to formally verify that the piece of code is constant time, but rely on actual measurements on the target platform to gather statistical evidence to disprove the null hypothesis "the code seems to run in constant time". Also, this is standard C code and you don't need any exotic tools to run it.
Warning: experimental quality software.
Oscar Reparaz (COSIC/KU Leuven)
(oscar dot reparaz at esat dot kuleuven dot be)