wei192026/passphrasegpt-mascara-attack

PassphraseGPT — Fine-tuned Attack Models

GPT-2 based passphrase language models fine-tuned for guessability research. Base model: PassphraseGPT (Kim et al., Open Sesame! On the Security and Memorability of Verbal Passwords).

Models Included

Directory	Fine-tuned on	Purpose
PassphraseGPT/last/	Original PassphraseGPT (no fine-tune)	Baseline attacker
PassphraseGPT_finetuned_user/last/	58,399 real-world leaked passphrases	Attack user-generated passphrases
PassphraseGPT_finetuned_markov/last/	500,000 Markov-chain passphrases	Attack Markov-generated passphrases
PassphraseGPT_finetuned_mascara/last/	500,000 MASCARA passphrases	Attack MASCARA-generated passphrases
PassphraseGPT_finetuned_diceware/last/	500,000 Diceware passphrases	Attack Diceware passphrases

Usage

Download and place under PassphraseGPT/pretrain/ in MASCARA-experiment:

hf download wei192026/passphrasegpt-mascara-attack \
    --local-dir PassphraseGPT/pretrain/ \
    --repo-type model
Then run Monte Carlo rank estimation:


python src/evaluate_passphrasegpt_mc.py \
    --model_path PassphraseGPT/pretrain/PassphraseGPT_finetuned_user/last \
    --tokenizer  PassphraseGPT/tokenizer/wordpiece/ \
    --test_set   corpus/user_test.txt \
    --n_samples  1000000 \
    --out_dir    results/ft_user_passphrasegpt

Citation

If you use these models, please cite:

@article{kim-open-sesame-verbal-passwords, author = {Eunsoo Kim and others}, title = {{Open Sesame! On the Security and Memorability of Verbal Passwords}}, }

@inproceedings{mukherjee-2023-memorable-passphrase, author = {Avirup Mukherjee and others}, title = {{MASCARA: Systematically Generating Memorable And Secure Passphrases}}, booktitle = {ACM ASIACCS 2023}, year = {2023} }