Adversarial ML Password Model

This repository contains the first draft of a trained LLM model that generates passwords from a word data set that contains the most used passwords in the world. We then modify those passwords and try to score a 3 or above on the zxcvbn strength meter test. After collecting 10k of these "strong" passwords. We train are TinyLLama LLM to make passwords that can fool password strength meters.

password_words.txt

Uses the most popular passwords in the world GitHub

generate_dataset.py

Creates the training passwords. Will create passwords till the 10k quota is met.

train_tinyllama.py

Tweaks/fine tunes the LLM.

generate_passwords_tinyllama.py

This is the main file that generates passwords with TinyLLama and runs it through zxcvbn.

evaluate.py

Will compare are dataset of passwords to are trained LLM passwords and display results.

Sequence of running files

Run in this order:

python3 generate_dataset.py

python3 train_tinyllama.py

python3 generate_passwords_tinyllama.py

python3 evaluate.py —generated (whatever the file name is)