A coarse-grained password model with memorable unit-based recurrent neural networks

Abstract

Passwords will continue to be the most prevalent form of authentication in the foreseeable future. But passwords often consist of some common segments which are easy to be predicted and attacked. Lots of methods have been proposed to describe password construction by regarding passwords as sequences of characters or segments. They are mainly based on statistical language models, for example Markov models and Probabilistic Context-Free Grammars. Recently, some researchers show that character-level recurrent neural networks can often outperform the previous methods. However, passwords usually consist of coarse-grained memorable units rather than fine-grained random characters. In this paper, we propose a memorable unit-based recurrent neural networks to model password construction. We first split passwords into memorable units (i.e., segments) with seven kinds of patterns and further split non-semantic pure letter and digit units into shorter units. Then, we build a unit-level recurrent neural network to predict the next unit with previous context units. Experiment results demonstrate that our coarse-grained unit-level model can reduce the search space during generating password guesses, which means our model obtains higher efficiency than the fine-grained character-level model.