Abstract
Most online services continue their reliance on text-based passwords as the primary authentication mechanism. With a growing number of these services and the limited creativity to devise new memorable passwords, users tend to reuse their passwords across multiple platforms. These factors, combined with the increasing number of leaked passwords, make passwords vulnerable to cross-site guessing attacks. Over the years, researchers have proposed several prevalent methods to predict subsequently used passwords, such as dictionary attacks, rule-based approaches, neural networks, and combinations of the above. We exploit the correlation between the similarity and predictability of these subsequent passwords in a dataset of 28.8 million users and their 61.5 million passwords. We use 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, twice as much as any other known model. The result makes our model effective for targeted online password guessing without getting spotted or locked out. To the best of our knowledge, this study is the first attempt of its kind using RNN. We also explore the use of RNN models in passphrase guessing. Passphrases are perceived to be more secure and easier to remember than passwords of the same length. We use a dataset that contains around 100,000 distinct phrases. We demonstrate that RNN models can predict complete passphrases given the initial word with rate up to 40%, which is twice better than other known approaches. Furthermore, our predictions can succeed in under 5,000 attempts, a 100% improvement compared to existing algorithms. In addition, this approach provides ease of deployment and low resource consumption. To our knowledge, it is the first attempt to exploit RNN for passphrase guessing.