PassFlow exploits the properties of Generative Flows to perform password guessing and shows very promising results, being competetive against GAN-based approaches [1, 2].
To get the dataset, run
curl -L --create-dirs -o data/train.txt https://github.com/d4ichi/PassGAN/releases/download/data/rockyou-train.txt
curl -L --create-dirs -o data/test.txt https://github.com/d4ichi/PassGAN/releases/download/data/rockyou-test.txt
and then run
pip install tqdm torch==1.7.1+cu110 -f https://download.pytorch.org/whl/torch_stable.html
to install the needed dependencies. We tested using PyTorch 1.7.1 and CUDA 11.0.
Once a model is trained, you can
python main.py --test <checkpoint>
to test the generation using the Static Sampling, or
python main.py --ds <checkpoint>
to test the generation using our Dynamic Sampling with Penalization, or
python main.py --gs <checkpoint>
to test the generation using our Dynamic Sampling with Gaussian Smoothing, as described in the paper.
[ 1 ] : https://arxiv.org/abs/1709.00440
[ 2 ] : https://arxiv.org/abs/1910.04232
PassFlow was made with ♥ and it is released under the MIT license.