Context Aware Password Guessability via Multi-Dimensional Rank Estimation

Password strength estimators are used to help users avoid picking weak passwords. Existing probabilistic estimators use various approaches to predict how many attempts a password cracker would need until it finds a given password. In this paper we present the first method for estimating the strength of human-chosen text passwords that is able to tweak the estimation according to each user's personal context, without retraining its model. Our method is able to incorporate context such as the user's name, preferred suffix, or previously cracked passwords (if available) when estimating the current password's strength. The tweaking takes only a few seconds per password. Our idea is to cast the question in a probabilistic framework used in side-channel cryptography. We view each password as a point in a d-dimensional search space, and learn the probability distribution of each dimension separately. The a-priori probability of a given password is the product of the d probabilities of its sub-passwords. After a detailed evaluation of leaked password corpora we found that an effective choice is to use d=5 dimensions: base word, prefix, suffix, shift-pattern, and l33t transformation. We coupled this decomposition with a state-of-the-art rank estimation algorithm to create our new PESrank estimator. We show that PESrank is more powerful than previous methods: it can crack more passwords, with fewer attempts, than the password crackers we compared it to. Even without using per-user context, PESrank is more accurate than previous methods for crackable passwords whose rank is smaller than 10^{12}. Furthermore, its training time is drastically shorter than previous methods, taking minutes, rather than days, to train on comparably-sized training sets, and taking a few hours to train on 905 million passwords, which is 8 times more passwords than previously used.