Using deep learning and NLP to analyze a large corpus of clear text passwords.
Objectives:
- Train a generative model.
- Understand how people change their passwords over time: hello123 -> h@llo123 -> h@llo!23.
Disclaimer: for research purposes only.
- 1.4 Billion Clear Text Credentials Discovered in a Single Database
- Collection of 1.4 Billion Plain-Text Leaked Passwords Found Circulating Online
- Archive of 1.4 BEEELLION credentials in clear text found in dark web archive
- Forbes
- Download any Torrent client.
- Here is a magnet link you can find on Reddit:
- magnet:?xt=urn:btih:7ffbcd8cee06aba2ce6561688cf68ce2addca0a3&dn=BreachCompilation&tr=udp%3A%2F%2Ftracker.openbittorrent.com%3A80&tr=udp%3A%2F%2Ftracker.leechers-paradise.org%3A6969&tr=udp%3A%2F%2Ftracker.coppersurfer.tk%3A6969&tr=udp%3A%2F%2Fglotorrents.pw%3A6969&tr=udp%3A%2F%2Ftracker.opentrackr.org%3A1337
- Checksum list is available here: checklist.chk
./count_total.sh
inBreachCompilation
should display something like 1,400,553,870 rows.
Process the data and run the first deep learning model:
# make sure to install the python deps first. Virtual env are recommended here.
# virtualenv -p python3 venv3; source venv3/bin/activate; pip install -r requirements.txt
# Remove "--max_num_files 100" to process the whole dataset (few hours and 50GB of free disk space are required.)
./process_and_train.sh <BreachCompilation path>
INPUT: BreachCompilation/
BreachCompilation is organized as:
- a/ - folder of emails starting with a
- a/a - file of emails starting with aa
- a/b
- a/d
- ...
- z/
- ...
- z/y
- z/z
OUTPUT: - BreachCompilationAnalysis/edit-distance/1.csv
- BreachCompilationAnalysis/edit-distance/2.csv
- BreachCompilationAnalysis/edit-distance/3.csv
[...]
> cat 1.csv
1 ||| samsung94 ||| samsung94@
1 ||| 040384alexej ||| 040384alexey
1 ||| HoiHalloDoeii14 ||| hoiHalloDoeii14
1 ||| hoiHalloDoeii14 ||| hoiHalloDoeii13
1 ||| hoiHalloDoeii13 ||| HoiHalloDoeii13
1 ||| 8znachnuu ||| 7znachnuu
EXPLANATION: edit-distance/ contains the passwords pairs sorted by edit distances.
1.csv contains all pairs with edit distance = 1 (exactly one addition, substitution or deletion).
2.csv => edit distance = 2, and so on.
- BreachCompilationAnalysis/reduce-passwords-on-similar-emails/99_per_user.json
- BreachCompilationAnalysis/reduce-passwords-on-similar-emails/9j_per_user.json
- BreachCompilationAnalysis/reduce-passwords-on-similar-emails/9a_per_user.json
[...]
> cat 96_per_user.json
{
"1.0": [
{
"edit_distance": [
0,
1
],
"email": "96-000@mail.ru",
"password": [
"090698d",
"090698D"
]
},
{
"edit_distance": [
0,
1
],
"email": "96-96.1996@mail.ru",
"password": [
"5555555555q",
"5555555555Q"
]
}
EXPLANATION: reduce-passwords-on-similar-emails/ contains files sorted by the first 2 letters of
the email address. For example 96-000@mail.ru will be located in 96_per_user.json
Each file lists all the passwords grouped by user and by edit distance.
For example, 96-000@mail.ru had 2 passwords: 090698d and 090698D. The edit distance between them is 1.
The edit_distance and the password arrays are of the same length, hence, a first 0 in the edit distance array.
Those files are useful to model how users change passwords over time.
We can't recover which one was the first password, but a shortest hamiltonian path algorithm is run
to detect the most probably password ordering for a user. For example:
hello => hello1 => hell@1 => hell@11 is the shortest path.
We assume that users are lazy by nature and that they prefer to change their password by the lowest number
of characters.
Run the data processing alone:
python3 run_data_processing.py --breach_compilation_folder <BreachCompilation path> --output_folder ~/BreachCompilationAnalysis
If the dataset is too big for you, you can set max_num_files
to something between 0 and 2000.
- Make sure you have enough free memory (8GB should be enough).
- It took 1h30m to run on a Intel(R) Core(TM) i7-6900K CPU @ 3.20GHz (on a single thread).
- Uncompressed output is around 45G.