A Study On A Small Sample Password Set Guessing Model Based On Multi-Task Learning

In recent years,with the increase of password data leakage and the enhancement of computer computing power,password guessing research has been greatly promoted.Password guess-ing analyzes password composition rules from a statistical point of view,studies password composition methods,security issues and usability,and proposes solutions for existing secu-rity issues.While improving password security,try to ensure the availability of passwords.However,most of the existing password guessing methods are driven by the data of a single password set,and the effect of guessing depends on the training data.Therefore,having a small password set leaked from a large password set,how to obtain a good guessing perfor-mance on the large password set is a key challenge.On the other hand,different password sets have different distribution,and it is difficult to use the information of other password sets to help guess another password set.Therefore,the difficulty of guessing a small sample password set lies in the small amount of data contained in the small sample set,and it is inconvenient to directly borrow password information from other password sets.From the perspective of learning the shared information of different password sets,this the-sis proposes a small sample password set guessing model based on multi-task learning to improve the guessing efficiency of small sample sets.The model extracts the password structure set,letter segment set,number segment set,and special character segment set from a password set,and uses the password structure sets of multiple password sets to pre-train a neural network in a multi-task learning manner.Then the neural network is trained again using the password structure set of the small sample set.Password structure set of the small sample set can be predicted using the trained neural network.The password guessing set can be obtained by filling the password structure set predicted by the neural network with the filling content obtained from the small sample set.This training method can use multiple password datasets for training,overcome the limitation of different distributions of different datasets,and fill in the missing password structure in the small sample set,thereby improving the guessing performance.The experiment uses three Chinese data sets and three English data sets respectively for empirical research.The results show that the password structure set generated by the small sample password set guessing model based on multi-task learning effectively expands the password structure set obtained directly from the small sample set.When the crack rate of PCFG＿v4.2 tends to be flat,the model proposed in this thesis can still ensure a continuous improvement of the crack rate.For example,taking small sample sets consisting of 2.6%and 0.21%of large sample sets respectively as the training set,after10~9guesses,the small sample password set guessing model based on multi-task learning can crack 21%password on average,which is 20%higher than PCFG＿v4.2.This thesis finds that an important factor affecting the improvement of the model’s cracking rate for a small sample password set is the size of the small sample set,and it is verified that a small sam-ple set containing 100,000 passwords already contains most of the password structure in the original large sample set,and 100,000 can be used as a dividing line to distinguish whether a password set is a small sample set.In addition,the elements that make up the password have higher transferability than the complete password.Further research on the transfer method of the letter segment and the number segment of the password can obtain better results in the small sample password guessing.