rarecoil / PassGAN

Python3, TensorFlow 1.15.0, CUDA 10.0 fork of PassGAN (https://arxiv.org/abs/1709.00440)

PassGAN

This is a fork of brannondorsey's implementation of PassGAN, a generative adversarial network built to make passwords. The repo was slightly out of date, relying on CUDA 8.0 and Tensorflow 1.4, as well as missing some dependencies that made it easy to train and run in 2019. View the original documentation at README.md.old for more information on this GAN.

Modifications

No modifications have been made to the underlying network implementation other than optimisation and refactoring for Python 3.x and TensorFlow 1.15.0, which makes it compatible with tensorflow-rocm and tensorflow-gpu.

Installation & Prerequisites

To install a CPU-only PassGAN and basic dependencies, run:

virtualenv venv
source venv/bin/activate
pip install -r requirements.txt

You will then need to install the correct GPU acceleration package. For ROCm, use:

pip install tensorflow-rocm==1.15.0

For NVIDIA/CUDA, use:

pip install tensorflow-gpu==1.15.0

Quick start

You can train your own model based upon your password corpus. The pretrained model has been deleted from this repository as it contains Python 2.x pickle data that is incompatible with Python 3. Training with large data sets will take quite some time with a reasonably fast GPU.

Training

From the original README.md.old:

# 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

Generation

This example uses a model you have trained to generate 1,000,000 passwords:

python sample.py \
	--input-dir $YOUR_OUTPUT \
	--checkpoint $YOUR_OUTPUT/checkpoints/$YOUR_CHECKPOINT.ckpt \
	--output gen_passwords.txt \
	--batch-size 1024 \
	--num-samples 1000000

Caveats

• This PassGAN implementation hardcodes the use of the NCHW format (for CuDNN) in a lot of places in tflib/, which is not supported on CPU (where it should be NHWC instead). Until this is patched to change the data format for CPU, you will need a GPU for training.
• Loading datasets for training will now assume that you are using UTF-8 for your dataset when running with Python 3. Note that many older dictionaries, such as rockyou.txt, are in ISO-8859-1. You can use iconv to transform them (e.g. iconv -f iso-8859-1 -t UTF-8 trainingdata/rockyou.txt > trainingdata/rockyou.utf8.txt).