Teaching a Neural Net How You Forget Your Passwords

What it does It ingests the BreachCompilation corpus of roughly 1.4 billion plaintext credentials and applies deep learning and NLP to treat passwords as a generative sequence. The pipeline clusters passwords by email address, measures edit distances between pairs, and attempts to reconstruct the chronological path of a user’s password mutations.

The interesting bit To guess the order in which a user changed passwords, the authors assume humans are lazy and prefer minimal edits. They run a shortest Hamiltonian path algorithm over a user’s password set to find the most probable evolutionary chain—turning credential drift into a graph problem driven by edit distance.

Key highlights

• Processes ~1.4 billion leaked credentials (torrent-based dataset, research-only disclaimer)
• Outputs edit-distance matrices and per-user password histories grouped by email prefix
• Infers likely password-change sequences via shortest-path graph algorithms
• Bundles a generative model training pipeline alongside the data-processing scripts
• Full processing demands about 8 GB of RAM, 50 GB of disk, and a few hours on a modern CPU

Caveats

• The README is thin on the actual deep-learning architecture; most documentation covers data preprocessing and graph inference rather than model internals
• You must source the BreachCompilation dataset yourself via the provided magnet link
• Running the full pipeline is resource-heavy and single-threaded

Verdict A solid starting point for security researchers studying credential lifecycle patterns, though the deep-learning portion feels like a black box next to the meticulous preprocessing. Skip it if you need a turnkey password auditor or a published reproducible paper.