This is a collection of password lists in which I have trained various deep learning algorithms to try to come up with passwords. A full report of results exists in each directory.
All models are trained against the top 10 million passwords in all of the hashes.org dataset, and then tested against the entire dataset for "cracks" vs. running the wordlists in hashcat.
- gpt2-small: Using minimaxir's gpt-2-simple with GPT2-small and some of the hashes.org "founds" dataset.
- PassGAN: Brannon Dorsey's implementation of the generative adversarial network used in PassGAN: A Deep Learning Approach for Password Guessing. The code is somewhat out of date; so I made a Python3/TensorFlow 1.15 fork.
This table lists approximate performance metrics for these models.
Metrics are:
- Model Name: The model/folder used to generate the password list.
- Generated: The amount of passwords generated by the model.
- Unique: The amount of unique passworss generated by the model.
- In Training Set: The amount of generated passwords that exist in the training dataset.
- PW/sec: Approximate password generation per second, raw.
- Cracks/sec.: Approximate novel passwords yielding cracks of the validation dataset per second.
- Rig: The computer specifications used for the experiment.
| Model Name | Generated | Unique | In Training Set | PW/sec. | Cracks/Sec. | Rig |
|---|---|---|---|---|---|---|
| gpt2-small | 4,726,912 | 4,053,784 | 822,690 | 111 | 12.17 | A1 |
| PassGAN | 4,999,168 | 4,658,237 | 454,759 | 84516 | 19726.01 | A2 |
This contains hardware specifications used to run the model.
- Intel Core i7-4790K (Devil's Canyon)
- 32GB DDR3 RAM
- 1 TB NVMe SSD (Samsung 960 Evo)
- Ubuntu 18.04.3 LTS, ROCm 2.9.6
- AMD Radeon VII (VBIOS version: 113-D3600200-106)
- Intel Core i7-4790K (Devil's Canyon)
- 32GB DDR3 RAM
- 1 TB NVMe SSD (Samsung 960 Evo)
- Ubuntu 18.04.3 LTS, CUDA 10.1
- NVidia Geforce GTX 1070 Ti Founders Edition
MIT License