Password Guessing Via Generative Adverarial Network

The identity authentication technology based on username-password is one of the most popular authentication methods due to its simplicity and flexibility.However,in order to relieve the stress on memory,users tend to create weak password or reuse a same password in different webs,resulting in severe security risk.Thus,how to improve the safety of password becomes an urgent problem to be solved,where the password guessing attack algorithms which have significant meanings in analyzing the composition of passwords and evaluating the strength of user passwords draw much research interest.During the past several decades,dozens of password guessing attack methods have been developed,such as constructing unique dictionaries,designing special guessing orders,and methods based on statistics and probabilistic,Markov-Chain and natural language processing.Although these methods have produced significant benefits,they have several drawbacks such as complicated processing and highly depending on the training sets.To address the above problems,in this thesis,we propose two password guessing attack algorithms based on the generate adversarial networks(GAN):1.Password guessing attack algorithms based on Text GAN.We utilize the recurrentneural network(RNN)and the convolutional neural network(CNN)as the generation model and discriminant model of GAN,respectively,where the gradient discreteproblem caused by the discrete data is solved by using an approximate function.2.Password guessing attack algorithms based on Seq GAN,which uses RNN as the generation model and discriminant model of GAN.Additionally,we use Monte-Carlotree search(MCTS)to fill up the incomplete sequences and introduce reinforcementlearning to guide the character generating.Through tremendous amount of experiments,we can show that the algorithm based on Seq GAN can crack the passwords more effectively and better compensate the shortcomings of the password guessing attack algorithm based on the conventional probabilistic model.Besides,the performance of the algorithm based on Text GAN reveals the limitation of the text generation model in the field of password cracking,which can produce certain guidance effect on the research of password guessing attack algorithm.