This repository contains code for the PassGAN: A Deep Learning Approach for Password Guessing paper.
The model from PassGAN is taken from Improved Training of Wasserstein GANs and it is assumed that the authors of PassGAN used the improved_wgan_training tensorflow implementation in their work. For this reason, I have modified that reference implementation in this repository to make it easy to train (train.py
) and sample (sample.py
) from. This repo contributes:
- A command-line interface
- A pretrained PassGAN model trained on the RockYou dataset
# requires CUDA to be pre-installed
pip install -r requirements.txt
Use the pretrained model to generate 1,000,000 passwords, saving them to gen_passwords.txt
.
python sample.py \
--input-dir pretrained \
--checkpoint pretrained/checkpoints/195000.ckpt \
--output gen_passwords.txt \
--batch-size 1024 \
--num-samples 1000000
Training a model on a large dataset (100MB+) can take several hours on a GTX 1080.
# download the rockyou training data
# contains 80% of the full rockyou passwords (with repeats)
# that are 10 characters or less
curl -L -o data/train.txt https://github.com/brannondorsey/PassGAN/releases/download/data/rockyou-train.txt
# train for 200000 iterations, saving checkpoints every 5000
# uses the default hyperparameters from the paper
python train.py --output-dir output --training-data data/train.txt
You are encouraged to train using your own password leaks and datasets. Some great places to find those include:
- LinkedIn leak (2.9GB, direct download)
- Exploit.in torrent (10GB+, 800 million accounts. Infamous!)
- Hashes.org: a shared password recovery site.
I've yet to do an exhaustive analysis of my attempt to reproduce the results from the PassGAN paper. However, using the pretrained rockyou model to generate 10โธ password samples I was able to match 630,347 (23.97%) unique passwords in the test data, using a 80%/20% train/test split.
In general, I am somewhat surprised (and dissapointed) that the authors of PassGAN referenced prior work in the ML password generation domain but did not compare their results to that research. My initial experience with PassGAN leads me to believe that it would significantly underperform both the RNN and Markov-based approaches mentioned in that paper and I hope that it is not for this reason that the authors have chosen not to compare results.
This code is released under an MIT License. You are free to use, modify, distribute, or sell it under those terms.
The majority of the credit for the code in this repository goes to @igul222 for his work on the improved_wgan_training. I've simply modularized his code a bit, added a command-line interface, and specialized it for the PassGAN paper.
The PassGAN research and paper was published by Briland Hitaj, Paolo Gasti, Giuseppe Ateniese, Fernando Perez-Cruz.