In order to improve the recognition accuracy of keystroke authentication, a methodology based on feature extraction of keystroke sequence is presented in this paper. Firstly, the data of the users’ keystroke feature information that has too much deviation with the mean deviation is filtered out. Secondly, the probability of each input key is calculated and 10 values which do not have the best features are selected. Thirdly, they are weighed and a score evaluating the extent to which the user could be authenticated successfully is calculated. The benefit of using a third-party data set is more objective and comparable. At last, the experimental result shows that the detector in this paper is more accurate than the other three. First, the data of the users’ keystroke feature information that has too much deviation with the mean deviation is filtered out. Secondly, the probability of each input key is calculated and 10 values ​​which do not have the best features are selected. Thirdly, they are weighed and a score evaluating the extent to which the user could be authenticated successfully is calculated. The benefit of using a third-party data set is more objective and comparable. At last, the experimental result shows that the detector in this paper is more accurate than the other three.