This is an implementation of Dumer's algorithm for Information Set Decoding.
It is written specifically to solve decoding challenges on this website.
This will not run if your CPU does not have the AVX2 instruction set. To check whether you have it or not you can run:
$ cat /proc/cpuinfo | grep avx2
Executable name is isd
.
- Syndrome decoding
$ wget http://decodingchallenge.org/Challenges/SD/SD_300_0
$ time ./isd 8 SD SD_300_0
n=300 k=150 w=38
l=16 p=4 epsilon=69 doom=0
001000010100000100000001000000000000000000000000000000001000000001010000000001000000000000000000000000000000000010000010000000000000000100000000000100100000000000011001000000000000000100000010000000000001100000010000000000000000000000001000101000000000010110001000010000010111000000000011010000000000
real 0m8.874s
user 1m8.215s
sys 0m0.156s
- Quasi-cyclic setting
$ wget http://decodingchallenge.org/Challenges/QC/Provider1/QC_28
$ time ./isd 8 QC QC_28
n=786 k=393 w=28
l=16 p=4 epsilon=40 doom=1
000000000000000000000000000000000000000100000000000000010010000010000000100010000000000000000000010100000100000000000000000000000000000000000000000000000000000000000000000000000000000001000000000000000000000000000000000000000000000000000000001000000000000000000000000000000000000000000010000000000000000000000000000000000000000000000000000000000000000000010000000000010000000001000000000000000000000010000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001000000000000000000000000000100000000000000000000000000000000000000010000010000000100000000000000000000000010000000000010000010000000000000000000000000000000000100000000000000000000000000000000000000100000000001010000000000000000000000000000000000000000000000000000000
real 0m3.661s
user 0m28.618s
sys 0m0.060s
Dumer parameters are chosen at compile time. They are:
DUMER_L
the width of the vectors used for collision in the birthday decoding partDUMER_P
the weight of the vectors searched using birthday decodingDUMER_EPS
the number of columns overlapping in the two setsDUMER_DOOM
set to 1 to use the quasi-cyclicity of a code (all the circular shift of a syndrome will give the same error pattern up to blockwise circular shifts)DUMER_LW
set to 1 to look for low-weight codeword instead of decoding
To set them the most convenient way is probably doing, for example:
$ cmake -B build -DDUMER_L=16 -DDUMER_P=4 -DDUMER_EPS=40 -DDUMER_DOOM=0 -DDUMER_LW=0 && cmake --build build/
A python script (optimize.py
) is provided to help choose the parameters of the algorithm.
It empirically estimates the average running time for fixed parameters.
The optimal parameters are chosen using a simple hill-climbing optimization of the running time as a function of the parameters.
Its arguments are the same as those for isd
.
The starting values for P
, L
, EPS
can be modified by appending, for example, P=5 L=30
.
If used for low-weight codeword finding, a target weight W
should be provided by appending it in the same fashion, for example, W=218
.
- Example in the quasi-cyclic setting
$ wget http://decodingchallenge.org/Challenges/QC/Provider0/QC_30
$ python3 scripts/optimize.py 8 QC QC_30 P=5
P L EPS Est. time
5 12 10 0:06:02.380958
5 12 9 0:06:06.140503
[...]
5 19 10 0:01:52.285594
5 19 9 0:01:53.148645
5 19 11 0:01:59.560413
5 20 10 0:01:58.913279
5 20 9 0:01:55.195710
5 20 11 0:01:53.693244
-DDUMER_P=5L -DDUMER_L=19L -DDUMER_EPS=10L
- Example in the low-weight codeword finding setting
$ wget https://decodingchallenge.org/Challenges/LW/LW_1280_0
$ python3 scripts/optimize.py 8 LW LW_1280_0 W=230
P L EPS Est. time
4 12 10 2:43:51.936259
4 12 9 2:39:50.257167
4 12 11 2:37:09.806838
[...]
4 12 12 2:37:03.774286
4 11 12 2:43:15.992123
4 13 12 2:48:15.027402
4 12 13 2:37:16.885877
4 11 13 2:47:01.252960
4 13 13 2:51:33.162128
-DDUMER_P=4L -DDUMER_L=12L -DDUMER_EPS=12L
Files should respect the formats of http://decodingchallenge.org/.
MIT (see file headers)