Learning Password Modification Patterns with Recurrent Neural Networks

The majority of online services continue their reliance on text-based passwords as the primary means of user authentication. With a growing number of these services and the limited creativity and memory to come up with new memorable passwords, users tend to reuse their passwords across multiple platforms. These factors, combined with the increasing amount of leaked passwords, make passwords vulnerable to cross-site guessing attacks. Over the years, several popular methods have been proposed to predict subsequently used passwords, such as dictionary attacks, rule-based approaches, neural networks, and combinations of the above. In this paper, we work with a dataset of 28.8 million users and their 61.5 million passwords, where there is at least one pair of passwords available for each user. We exploit the correlation between the similarity and predictability of these subsequent passwords. We build on the idea of a rule-based approach but delegate rule derivation, classification, and prediction to a Recurrent Neural Network (RNN). We limit the number of guessing attempts to ten yet get an astonishingly high prediction accuracy of up to 83% in under five attempts in several categories, which is twice as much as any other known models or algorithms. It makes our model an effective solution for real-time password guessing against online services without getting spotted or locked out. To the best of our knowledge, this study is the first attempt of its kind using RNN.