Skip to content
A scalable search index for binary files
C++ M4 Python Makefile Shell
Branch: master
Clone or download

Latest commit

Fetching latest commit…
Cannot retrieve the latest commit at this time.


Type Name Latest commit message Commit time
Failed to load latest commit information.


BigGrep is a tool to index and search a large corpus of binary files that uses a probabalistic N-gram based approach to balance index size and search speed.


BigGrep requires Boost version 1.48 or later. Boost should be available in most package management systems. Build and install Boost before building BigGrep. To use the Boost lockfree queue, version 1.53 or greater should be installed. This may or may not give you a performance boost when indexing.

bgsearch requires jobdispatch, a python package that is included and installed automatically with biggrep

git clone  
cd BigGrep  
make install

Now let's make a couple of test indexes out of some Windows EXE files:

mkdir /tmp/bgi  
ls -1 /some/test/files/*.exe | bgindex -p /tmp/bgi/testidx1 -v  
ls -1 /some/more/test/files/*.exe | bgindex -p /tmp/bgi/testidx2 -v  

And now that we have executables and test indexes, here's some sample search usage with verification (in this case, searching for a typical function entry point byte sequence, not overly interesting but shows how simple it is to look for the existence of an abitrary byte sequence) using bgsearch:

bgsearch -d /tmp/bgi/ -v 8bff558bec  
bgsearch -d /tmp/bgi/ -v program  

More Info

For additional details, see the original paper (also found in this repository in the doc directory):

Jin, W.; Hines, C.; Cohen, C.; Narasimhan, P., "A scalable search index for binary files," Malicious and Unwanted Software (MALWARE), 2012 7th International Conference on , vol., no., pp.94,103, 16-18 Oct. 2012
doi: 10.1109/MALWARE.2012.6461014

There is also a whitepaper in the doc directory of this repo that was done a little later and describes more implementation decisions and some of our real world usage details at the time (including some things that changed after the paper for MALWARE 2012 was written based on things discovered during our daily usage, such as switching to 3-grams for the indexes to save on disk space).

Later Changes

More minor changes to the code and our usage have also occurred since that whitepaper was written, such as:

  • The current "hints" list in the indexes was modified to get us within 16 N-grams instead of 256 to help further reduce I/O at the expense of slightly increasing the index size (this is configurable at index generation time)
  • We now have gone to a mixed collection of 3-gram and 4-gram indexes to help eliminate some more I/O issues while continuing to balance disk space and speed. Basically, for a small percentage of the files (just the ones that fill up a large portion of the 3-gram space, making them frequently appear as false positives in the candidate searches with 3-gram indexes) we generate 4-gram indexes and use 3-grams for the rest of them. By using the -O and -m options in bgindex, a file can be rejected from a 3-gram index if it fills a large portion of the 3-gram space. By using the -O option, bgindex will write the file path to the file given to -O and you can then generate 4-gram indexes for the denser files.
  • The current search code can also filter on metadata values to further trim down the candidate list before verification. The man pages give details on that.

Major Components

A quick overview of the various components (see the docs and source for more info):

  • bgindex: to create an index from a list of files
  • bgsearch: Python wrapper to search the indexes for the desired byte sequence(s) and optionally invoke the verification step and filter on metadata embedded in the indexes. It tries to guess what you are searching for by inspecting the seach strings (if you don't explicitly tell it via a cmd line option): hex byte values or an ASCII string (which it converts to hex byte values to do the search). Can use bgverify or Yara (with a supplied rule file) to do the verification.
  • bgparse: the executable bgsearch calls that actually reads the indexes to do the candidate list for the files with N-grams of the byte values.
  • bgverify: a Boyer-Moore-Horspool fast string search based verification tool to make sure the full strings of byte values exist in the candidate files found by bgparse. Good for simple verifications, but as mentioned above you can also use Yara instead (see bgsearch docs).
  • bgextractfile: removes or replaces a file from an index. This is useful if files have been purged or moved but you don't want to re-index.


A couple of minor notes about the code & building it:

  • If installing from the tarball, you should not have to run ./ ./configure; make; make install should configure, build, and install this package.

  • The code depends on Boost and Python (2.6 or 2.7, untested with 3.x), and is known to build on various versions of RedHat, CentOS, MAC OSX and Ubuntu Linux distros.

  • By default, the biggrep and jobdispatch python packages are installed in the site-packages directory returned by disutils.sysconfig get_python_lib(). You can change this behavior by using --with-python-prefix{=DIR} to install the Python modules under this prefix (location will be DIR/lib/python*/biggrep). If DIR is not provided, it will use the value of PREFIX (--prefix).

  • If using a prefix when installing BigGrep, you may need to use --with-python-prefix to install the required Python modules.

  • If using --with-python-prefix you may need to set your PYTHONPATH environment variable to the location where the biggrep python module was installed in order to run bgsearch.

  • Boost 1.48 is required. On RHEL6 you may be able to find this package in the Software Collections Library, however, it may install in an alternative path. Try the following when linking against the boost148 package:

    LDFLAGS=-L/usr/lib64/boost148 LIBS="-lboost_system-mt -lboost_chrono-mt" ./configure BOOST_ROOT=/usr/include
    make install


Feel free to write an Issue for any problems you encounter.

Copyright 2011-2017 Carnegie Mellon University. See LICENSE file for terms.


You can’t perform that action at this time.