No description, website, or topics provided.
C++ Java Shell CMake JavaScript Python Other
Clone or download
jonjoliver Merge pull request #53 from EdgewiseNetworks/master
CMake changes to build on Windows
Latest commit 7617ad5 Apr 15, 2018

Build Status

TLSH - Trend Micro Locality Sensitive Hash

TLSH is a fuzzy matching library. Given a byte stream with a minimum length of 50 bytes TLSH generates a hash value which can be used for similarity comparisons. Similar objects will have similar hash values which allows for the detection of similar objects by comparing their hash values. Note that the byte stream should have a sufficient amount of complexity. For example, a byte stream of identical bytes will not generate a hash value.

Minimum byte stream length

The program in default mode requires an input byte stream with a minimum length of 256 bytes (and a minimum amount of randomness - see note in Python extension below).

In addition the -force option allows strings down to length 50. See notes for version 3.5.0 of TLSH

Computed hash

The computed hash is 35 bytes long (output as 70 hexidecimal charactes). The first 3 bytes are used to capture the information about the file as a whole (length, ...), while the last 32 bytes are used to capture information about incremental parts of the file. (Note that the length of the hash can be increased by changing build parameters described below in CMakeLists.txt, which will increase the information stored in the hash. For some applications this might increase the accuracy in predicting similarities between files.)

Executables and library

Building TLSH (see below) will create a static library in the lib directory, and the tlsh executable (a symbolic link to tlsh_unittest). 'tlsh' links to the static library, in the bin directory. The library has functionality to generate the hash value from a given file, and to compute the similarity between two hash values.

tlsh is a utility for generating TLSH hash values and comparing TLSH hash values to determine similarity. Run it with no parameters for detailed usage.


  • A JavaScript port available in the js_ext directory.
  • A Java port is available in the java directory.
  • Another Java port is available here.
  • A Golang port is available here.
  • A Ruby port is available here

Downloading TLSH

Download TLSH as follows:

wget -O
cd tlsh-master


git clone git://
cd tlsh
git checkout master

Building TLSH

Edit CMakeLists.txt to build TLSH with different options.

  • TLSH_BUCKETS: determines using 128 or 256 buckets use the default 128 buckets unless you are an expert and know you need 256 buckets
  • TLSH_CHECKSUM_1B: determines checksum length, longer means less collision use the default 1 byte unless you are an expert and know you need a larger checksum



Note: Building TLSH on Linux depends upon cmake to create the Makefile and then make the project, so the build will fail if cmake is not installed.

Windows (Visual Studio)

Use the version-specific tlsh solution files (tlsh.VC2005.sln, tlsh.VC2008.sln, ...) under the Windows directory.

See tlsh.h for the tlsh library interface and tlsh_unittest.cpp and simple_unittest.cpp under the test directory for example code.

Python Extension

There is a README.python with notes about the python version

(1) compile the C++ code
(2) build the python version
	$ cd py_ext/
	$ python ./ build
(3) install - possibly - sudo, run as root or administrator
	$ python ./ install
(4) test it
	$ cd ../Testing
	$ ./

Python API

import tlsh

Note that in default mode the data must contain at least 256 bytes to generate a hash value and that it must have a certain amount of randomness. If you use the "force" option, then the data must contain at least 50 characters. For example, tlsh.hash(str(os.urandom(256))), should always generate a hash.
To get the hash value of a file, try tlsh.hash(open(file, 'rb').read()).

tlsh.diff(h1, h2)
tlsh.diffxlen(h1, h2)

The diffxlen function removes the file length component of the tlsh header from the comparison. If a file with a repeating pattern is compared to a file with only a single instance of the pattern, then the difference will be increased if the file lenght is included. But by using the diffxlen function, the file length will be removed from consideration.

Note that the python API has been extended to miror the C++ API. See py_ext/tlshmodule.cpp and the py_ext/ script to see the full API set.

Design Choices

  • To improve comparison accuracy, TLSH tracks counting bucket height distribution in quartiles. Bigger quartile difference results in higher difference score.
  • Use specially 6 trigrams to give equal representation of the bytes in the 5 byte sliding window which produces improved results.
  • Pearson hash is used to distribute the trigram counts to the counting buckets.
  • The global similarity score distances objects with significant size difference. Global similarity can be disabled. It also distances objects with different quartile distributions.
  • TLSH can be compiled to generate 70 or 134 characters hash strings. The longer version has been created to use of the 70 char hash strings is not working for your application.

TLSH similarity is expressed as a difference score:

  • A score of 0 means the objects are almost identical.
  • For the 70 characters hash, there is a detailed table of experimental Detection rates and False Positive rates based on the threshhold. see Table II on page 5




  • Implemented TLSH.
  • Updated to build with CMake.


  • Enabled C++ optimization. Runs 4x faster.


  • Supports Windows and Visual Studio.


  • Added Python extension library. TLSH is callable in Python.
  • Stop generating hash if the input is less than 512 bytes.
  • Cleaned up.


  • Length difference consideration can be disabled in this version. See totalDiff in tlsh.h.
  • TLSH can be compiled to generate the 70 or 134 character hashes. The longer version is more accurate.


  • The checksum can be changed from 1 byte to 3 bytes. The collison rate is lower using 3 bytes.
  • If the incoming data has few features. The algorithm will not generate hash value. At least half the buckets must be non-zero.
  • Null or invalid hash strings comparison will return -EINVAL (-22).
  • Python extension library will read CMakeLists.txt to pick the compile options.
  • The default build will use half the buckets and 1 byte checksum.
  • New executable tlsh_version reports number of buckets, checksum length.


  • Add and scripts for building/cleaning the project.
  • Modifications to tlsh_unittest.cpp to write errors to stderr (not stdout) and to continue processing in some error cases. Also handle a listfile (-l parameter) which contains both TLSH and filename.
  • Updated expected output files based on changes to tlsh_unittest.cpp.


  • Updated the Testing/exp expected results.
  • Created a script to ease the creation of the Testing/exp expected results.


  • Updated tlsh_util.h, tlsh_impl.cpp, tlsh_util.cpp on checksum.
  • Updated and Testing/exp results.


  • Add Visual Studio 2005 and 2008 project and solution files to enable build on Windows environment.
  • Added files WinFunctions.h and WinFunctions.cpp to handle code changes needed for Windows build.
  • Modified several unit test expected output files to remove error messages, to allow the running of unit tests on Windows under Cygwin. This was caused by the opposite order in which stdout and stderr are written when stderr is redirected to stdout as 2>&1. Also modified to write stderr to /dev/null.
  • Move rand_tags executable from tlsh_forest project to tlsh, to reduce the dependencies of the tlsh ROC analysis project, which depends upon tlsh_unittest and rand_tags.
  • Remove simple_unittest and tlsh_version from bin directory as these executables are for internal testing and source code documentation, and do not need to be exported.
  • Add -version flag to tlsh_unittest to get the version of the tlsh library.


  • Pickup fix to hash_py() in py_ext/tlshmodule.cpp (commit da5370bcfdd40dd6a33c877ee87fe3866188cf2d).


  • Made the minimum data length = 256 for the C version. This was reduced in version 3.5.0 to 50 bytes (the force option)


  • Fixed bug introduced by commit 1a8f1c581c8b988ced683ff8e0a0f9c574058df4 which caused a different hash value to be generated if there were multiple calls to Tlsh::update as opposed to a single call.


  • Add JavaScript implementation (see directory js_ext) - required for Blackhat presentation.
  • Modify tlsh_unittest so that it can output tlsh values and filenames correctly, when the filenames contain embedded newline, linefeed or tab characters.


  • Thanks to Jeremy Bobbios py_ext patch. TLSH has these enhancements.
  • Instead of using a big memory blob, it will calculate the hash incrementally.
  • A hashlib like object-oriented interface has been added to the Python module. See
  • Restrict the function to be fed bytes-like object to remove surprises like silent UTF-8 decoding.


  • Back out python regression test as part of the script, so that the python module does not need to be installed in order to successfully pass the tests run by


  • Fix regression tests running on Windows


  • Specify Tlsh::getHash() is a const method


  • Pick up Jeremy Bobbios patches for:
    • Build shared library (, in addition to static library, on Linux and have tlsh_unittest link to it.
    • Remove TlshImpl symbols from
    • Add Tlsh_init to py_ext/tlshmodule.cpp, which ensures Tlsh constructor will be called from Tlsh python module
    • Create symbolic link for tlsh -> tlsh_unittest


  • Added the - force option
    • Allows a user to force the generation of digests for strings down to 50 characters long


  • Fixed the error in the Python extension


  • Added the BlackHat Asia tool (presented at Arsenal)


  • skipped


  • merged in various fixes - ifdef for SPARC and RH73 corrected TLSH_CTC_final.pdf (see added a SHA1 to the NOTICE.txt file improved the so that it calls the (and does regression tests) improved regression tests to confirm that the hash is calculated correctly in your environment fixed the header file C++ standard violation (reserved identifier violation #21)


  • resolved issue #29 - the force option for Python Step 1 - adding a regression test for strings approx of length 50 Step 2 - add python code


  • added code to set the distance parameters for ROC analysis to use these settings then change in CMakeLists.txt set(TLSH_DISTANCE_PARAMETERS 0) => set(TLSH_DISTANCE_PARAMETERS 1)


  • resolving issue #44
  • making static library the default


  • resolving issue #45
  • add a timing test for TLSH
  $ bin/timing_unittest
  build a buffer with a million bytes...
  eval TLSH 50 times...
  TLSH(buffer) = A12500088C838B0A0F0EC3C0ACAB82F3B8228B0308CFA302338C0F0AE2C24F28000008
  BEFORE	ms=1502905567631
  AFTER	ms=1502905573523
  TIME	ms=5892
  TIME	ms=117	per iteration


  • resolving issue #46
  • in include/tlsh_impl.h #define SLIDING_WND_SIZE 5 this can be varied between 4 to 8

3.8.0 Adding // access functions - required by tools using TLSH library

  • int Lvalue();
  • int Q1ratio();
  • int Q2ratio();

3.9.0 resolving issue #48 - tlsh_pattern program This tlsh_pattern program should read a pattern file

  • col 1: pattern number
  • col 2: nitems in group
  • col 3: TLSH
  • col 4: radius
  • col 5: pattern label The input options should match the tlsh program
usage: tlsh_pattern [-xlen] [-force] -pat pattern_file -f file
: tlsh_pattern [-xlen] [-force] -pat pattern_file -d digest
: tlsh_pattern [-xlen] [-force] -pat pattern_file -r dir
: tlsh_pattern [-xlen] [-force] -pat pattern_file -l listfile

3.9.1 resolving issue #38 putting in fix in rand_tags.cpp so that it generates identical output to previous version while safely working with pointers