This provides a class that performs symbolic execution of MT19937, as well as a solver for GF(2) matrices and precomputed solutions for certain instances of MT19937 cloning. The solver has a Python-Only solver with no other dependencies and a python wrapper for Cryptominisat to solve GF(2) matrices. Note that Cryptominisat has to be built with GAUSS.
The Python-Only solver is faster than Cryptominisat built without M4RI but takes up a lot of RAM.
A demo of all the main features are present in the python notebook Demo of Features.ipynb
There are three ways to clone MT19937 provided here. The first is using one of the precomputed solutions, which are used if the known numbers are the 4*k most significant bits of consecutive outputs of an MT19937 generator, e.g python's random.getrandbits(nbits). This usually solves in 1-2 seconds. The next two are with the Python-Only solver and Cryptominisat. The Python-Only solver takes around a minute for sparse matrices but up to two hours for dense matrices. I did not do tests with Cryptominisat as I wasn't able to build Cryptominisat with M4RI on Windows. I suggest building Cryptominisat with M4RI for the speed. Regardless, a Windows x64 built of Cryptominisat with GAUSS but no M4RI is included in the bin folder.
This also provides a way to reverse the state of an MT19937 generator, to predict numbers generated before the known numbers. However, as of now there is a bug somewhere that makes it work only sometimes. In the event that reversing does not work, simply forward the cloned MT19937 and reverse.
These are all demonstrated in Demo of Features.ipynb
numpy
gzip
Do note to use the Cryptominisat Python wrapper you are required to first build Cryptominisat with GAUSS. For speed it should also be built with M4RI.