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.
What's New in TLSH 4.10.x
Release version 4.8.x - merged in pull requests for more stable installation.
Release version 4.9.x - added -thread and -private options.
- Both versions are faster than previous versions, but they set the checksum to 00.
- This loses a very small part of the functionality.
- See 4.9.3 in the Change_History to see timing comparisons.
Release version 4.10.x - a Python clustering tool.
- See the directory tlshCluster.
TLSH has gained some traction. It has been included in STIX 2.1 and been ported to a number of langauges.
We have added a version identifier ("T1") to the start of the digest so that we can cleary distinguish between different variants of the digest (such as non-standard choices of 3 byte checksum). This means that we do not rely on the length of the hex string to determine if a hex string is a TLSH digest (this is a brittle method for identifying TLSH digests). We are doing this to enable compatibility, especially backwards compatibility of the TLSH approach.
The code is backwards compatible, it can still read and interpret 70 hex character strings as TLSH digests. And data sets can include mixes of the old and new digests. If you need old style TLSH digests to be outputted, then use the command line option '-old'
Thanks to Chun Cheng, who was a humble and talented engineer.
Minimum byte stream length
The program in default mode requires an input byte stream with a minimum length of 50 bytes (and a minimum amount of randomness - see note in Python extension below).
For consistency with older versions, there is a -conservative option which enforces a 256 byte limit. See notes for version 3.17.0 of TLSH
The computed hash is 35 bytes of data (output as 'T1' followed 70 hexidecimal characters. Total length 72 characters). The 'T1' has been added as a version number for the hash - so that we can adapt the algorithm and still maintain backwards compatibility. To get the old style 70 hex hashes, use the -old command line option.
Bytes 3,4,5 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
tlsh executable (a symbolic link to
'tlsh' links to the static library, in the
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 Java port is available in the
3rd Party Ports
We list these ports just for reference. We have not checked the code in these repositories, and we have not checked that the results are identical to TLSH here. We also request that any ports include the files LICENSE and NOTICE.txt exactly as they appear in this repository.
- Another Java port is available here.
- Another Java port is available here.
- A Golang port is available here.
- A Ruby port is available here
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
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.
To install cmake/gcc compiler on CentOs or Amazon Linux:
$ sudo yum install cmake
$ sudo yum install gcc-c++
Added in March 2020. See the instructions in README.mingw
Windows (Visual Studio)
Using TLSH in Python
$ pip install py-tlsh
If you need to build your own Python package, then there is a README.python with notes about the python version
(1) compile the C++ code $./make.sh (2) build the python version $ cd py_ext/ $ python ./setup.py build (3) install - possibly - sudo, run as root or administrator $ python ./setup.py install (4) test it $ cd ../Testing $ ./python_test.sh
import tlsh tlsh.hash(data)
Note data needs to be bytes - not a string. This is because TLSH is for binary data and binary data can contain a NULL (zero) byte.
In default mode the data must contain at least 50 bytes to generate a hash value and that it must have a certain amount of randomness. To get the hash value of a file, try
Note: the open statement has opened the file in binary mode.
import tlsh h1 = tlsh.hash(data) h2 = tlsh.hash(similar_data) score = tlsh.diff(h1, h2) h3 = tlsh.Tlsh() with open('file', 'rb') as f: for buf in iter(lambda: f.read(512), b''): h3.update(buf) h3.final() # this assertion is stating that the distance between a TLSH and itself must be zero assert h3.diff(h3) == 0 score = h3.diff(h1)
Python Extra Options
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.
Python Backwards Compatibility Options
If you use the "conservative" option, then the data must contain at least 256 characters. For example,
import os tlsh.conservativehash(os.urandom(256))
should generate a hash, but
will generate TNULL as it is less than 256 bytes.
If you need to generate old style hashes (without the "T1" prefix) then use
The old and conservative options may be combined:
- 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 72 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
- See the Python code and Jupyter notebooks in tlshCluster.
- We provide Python code for the HAC-T method. We also provide code so that users can use DBSCAN.
- We show users how to create dendograms for files, which are a useful diagram showing relationships between files and groups.
- We provide tools for clustering the Malware Bazaar dataset, which contains a few hundred thousand samples.
- The HAC-T method is described in HAC-T and fast search for similarity in security
- 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
- Jonathan Oliver, Muqeet Ali, and Josiah Hagen. HAC-T and fast search for similarity in security 2020 International Conference on Omni-layer Intelligent Systems (COINS). IEEE, 2020.
23/10/2021 resolve issue #116 Library will not compile on CENTOS 7 (use of threads)