Password Guessing Based on Semantic Analysis and Neural Networks

Passwords remain the dominant method in data encryption and identity authentication, but they are vulnerable to guessing attack. Most users incline to choose meaningful words to make up passwords. Lots of these words are human-memorable. In this paper, we propose a hierarchical semantic model that combines LSTM with semantic analysis to implement password guessing. With our model, the potential probability relationship between words can be mined. After training the model with 4.5 million passwords from leaked Chinese passwords, we generate lots of passwords guesses ordered by probability. 0.5 million passwords are reserved for model testing. In addition, we also pick up CSDN passwords, the Rockyou passwords, and Facebook passwords as model-testing sets. Each dataset contains 0.5 million passwords. LSTM-based model, PCFG, and Markov-based model are selected for comparison. Experiments show that our model has a higher coverage rate than the other models of the reserved dataset and CSDN dataset. Besides, our model can hit more passwords for the Rockyou dataset and Facebook dataset than PCFG.