TLSH - Trend Micro Locality Sensitive Hash

TLSH is a fuzzy matching library. Given a byte stream with a minimum length of 256 bytes (and a minimum amount of randomness - see note in Python extension below), 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.

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, and thus its accuracy in predicting similarities between files.)

Building TLSH (see below) will create a static library in the lib directory, and the tlsh_unittest executable, which 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_unittest is a utility for generating TLSH hash values and comparing TLSH hash values to determine similarity. Run it with no parameters for detailed usage.

Ports

A Java port is available here.
A JavaScript port available in the js_ext directory.

Downloading TLSH

Download TLSH as follows:

wget https://github.com/trendmicro/tlsh/archive/master.zip -O master.zip
unzip master.zip
cd tlsh-master

git clone git://github.com/trendmicro/tlsh.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, more is better
TLSH_CHECKSUM_1B: determines checksum length, longer means less collision

Linux

Execute:

make.sh

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

cd py_ext
python setup.py build
python setup.py install (sudo, run as root or administrator)

Python API

import tlsh
tlsh.hash(data)

Note that the data must contain at least 256 bytes to generate a hash value and that it must have a certain amount of randomness. 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/test.py 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 is more accurate.

TLSH similarity is expressed as a difference score:

A score of 0 means the objects are almost identical.
For the 70 characters hash, a score of 200 or higher means the objects are very different. For the 134 characters hash, a score of 400 or higher means the objects are very different.

Publications

Jonathan Oliver, Chun Cheng, and Yanggui Chen, TLSH - A Locality Sensitive Hash. 4th Cybercrime and Trustworthy Computing Workshop, Sydney, November 2013
Jonathan Oliver, Scott Forman, and Chun Cheng, Using Randomization to Attack Similarity Digests. ATIS 2014, November, 2014, pages 199-210

Changes

3.0.0

Implemented TLSH.
Updated to build with CMake.

3.0.1

Enabled C++ optimization. Runs 4x faster.

3.0.2

Supports Windows and Visual Studio.

3.0.3

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

3.0.4

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.

3.1.0

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.

3.1.1

Add make.sh and clean.sh 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.

3.1.2

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

3.1.3

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

3.2.0

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 test.sh 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.

3.2.1

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

3.3.0

Made the minimum data length = 256 for the C version.

3.3.1

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.

3.4.0

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.

3.4.1

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 test.py.
Restrict the function to be fed bytes-like object to remove surprises like silent UTF-8 decoding.

3.4.2

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

3.4.3

Fix regression tests running on Windows

3.4.4

Specify Tlsh::getHash() is a const method

3.4.5

Pick up Jeremy Bobbios patches for:
- Build shared library (libtlsh.so), in addition to static library, on Linux and have tlsh_unittest link to it.
- Remove TlshImpl symbols from libtlsh.so
- 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

mknjc/tlsh