yujiaxiao/passrvae

PassRVAE

This repo covers the implementation for our paper PassRVAE.

Environments

Python 3.10.12
Pytorch 2.0.1+cu118

Datasets

In our experiments, datasets include RockYou and 4iQ. To get passwords of different composition policies (1class8, 2class8, 3class8, 4class8, 3class12 and 1class16), we use the following regular expressions.

passwords with at least one category class (uppercase letters, lowercase letters, symbols, digits)

grep -a -E '[[:upper:]]|[[:lower:]]|[[:digit:]]|[[:punct:]]'

passwords with at least two category classes (uppercase letters, lowercase letters, symbols, digits)

grep -a -E '([[:upper:]].[[:lower:]])|([[:upper:]].[[:punct:]])|([[:upper:]].[[:digit:]])|([[:digit:]].[[:punct:]])|([[:lower:]].[[:digit:]])|([[:lower:]].[[:punct:]])'

passwords with at least three category classes (uppercase letters, lowercase letters, symbols, digits)

grep -a -E '([[:upper:]].[[:lower:]].[[:digit:]])|([[:upper:]].[[:lower:]].[[:punct:]])|([[:upper:]].[[:digit:]].[[:punct:]])|([[:lower:]].[[:digit:]].[[:punct:]])'

passwords with at least four category classes (uppercase letters, lowercase letters, symbols, digits)

grep -a -E '[[:upper:]].[[:lower:]].[[:digit:]].*[[:punct:]]'

In the following steps, we choose the 20:80 split of 4class8 of 4iQ as an example.

Download code

git clone https://github.com/yujiaxiao/PassRVAE PassRVAE
cd PassRVAE

Train

python3 train.py

Output:
TRAIN preprocessed file not found at ./data/4iq-4class8-20%-train.json. Creating new.
Vocablurary of 114 keys created.
VALID preprocessed file not found at ./data/4iq-4class8-20%-valid.json. Creating new.
preprocess data time 15.288792371749878
PassRVAE(
(embedding): Embedding(114, 300)
(embedding_dropout): Dropout(p=0.1, inplace=False)
(encoder_rnn): GRU(300, 256, num_layers=3, batch_first=True)
(decoder_rnn): GRU(300, 256, num_layers=3, batch_first=True)
(hidden2mean): Linear(in_features=768, out_features=128, bias=True)
(hidden2logv): Linear(in_features=768, out_features=128, bias=True)
(latent2hidden): Linear(in_features=128, out_features=768, bias=True)
(outputs2vocab): Linear(in_features=256, out_features=114, bias=True)
)
TRAIN Epoch 00/20, Mean ELBO 28.5835
Model saved at ./bin/2023-Oct-02-12:56:35/E0.pytorch
VALID Epoch 00/20, Mean ELBO 35.1942
TRAIN Epoch 01/20, Mean ELBO 33.9293
Model saved at ./bin/2023-Oct-02-12:56:35/E1.pytorch
VALID Epoch 01/20, Mean ELBO 33.0856
TRAIN Epoch 02/20, Mean ELBO 32.6987
Model saved at ./bin/2023-Oct-02-12:56:35/E2.pytorch
VALID Epoch 02/20, Mean ELBO 32.5687
TRAIN Epoch 03/20, Mean ELBO 32.2161
Model saved at ./bin/2023-Oct-02-12:56:35/E3.pytorch
VALID Epoch 03/20, Mean ELBO 32.3595
TRAIN Epoch 04/20, Mean ELBO 31.9364
Model saved at ./bin/2023-Oct-02-12:56:35/E4.pytorch
VALID Epoch 04/20, Mean ELBO 32.2145
TRAIN Epoch 05/20, Mean ELBO 31.7499
Model saved at ./bin/2023-Oct-02-12:56:35/E5.pytorch
VALID Epoch 05/20, Mean ELBO 32.1876
TRAIN Epoch 06/20, Mean ELBO 31.6235
Model saved at ./bin/2023-Oct-02-12:56:35/E6.pytorch
VALID Epoch 06/20, Mean ELBO 32.1332
TRAIN Epoch 07/20, Mean ELBO 31.5345
Model saved at ./bin/2023-Oct-02-12:56:35/E7.pytorch
VALID Epoch 07/20, Mean ELBO 32.0708
TRAIN Epoch 08/20, Mean ELBO 31.4716
Model saved at ./bin/2023-Oct-02-12:56:35/E8.pytorch
VALID Epoch 08/20, Mean ELBO 32.0599
TRAIN Epoch 09/20, Mean ELBO 31.4336
Model saved at ./bin/2023-Oct-02-12:56:35/E9.pytorch
VALID Epoch 09/20, Mean ELBO 32.0248
TRAIN Epoch 10/20, Mean ELBO 31.3976
Model saved at ./bin/2023-Oct-02-12:56:35/E10.pytorch
VALID Epoch 10/20, Mean ELBO 32.0790
Early stopping

Generate

Now you can use your trained model in the previous step to generate passwords. Or you can directly use our pretrained model to generate passwords.

python3 generate.py

Output:
Model loaded from ./pretrained/PassRVAE_4class8_20%.pytorch