The Research Of Keystroke Dynamics Based On Statistical Method

As the first security barrier of information system, the importance of identity authentication is self-evident. With the development of networking, the increased computation capacity of computer, and the enhanced awareness of copyright, traditional authentication based on username/password pair reflects many disadvantages. Common used biometrics, such as finger identification, iris recognition, etc., can overcome the flaws to a large extent, but it also brings some new problems, such as the higher cost, the complexity of deployment, and low level of user acceptance, and so on. As one of the biometrics, keystroke dynamics not only has the general advantages of biometrics, but also overcomes their disadvantages. Therefore, the application of keystroke dynamics is promising.In this paper, keystroke dynamics based on statistical method was studied. The model of keystroke dynamics authentication was built with HMM; the parameters in the model were computed by using Gaussian distribution, maximum likelihood estimation and other statistical knowledge; the evaluation problem was solved by using modified forward algorithm. At last, it could authenticate the identity of users by using the built model. As one of the behavioral characteristics, keystroke pattern changes through time. To solve this problem and improve the system performance, the idea of keystroke profile updating was proposed, and three updating strategy, namely full updating of the profile, part updating of the profile, and updating based on threshold (the combination of full updating and part updating) were proposed. At the same time, for the purpose of reducing the impact of input speed from the same user, numerical normalization of the keystroke time feature data was proposed. At the end of the paper, some experiments were did, and the results confirmed that it can improve the system performance (reduce the AFR (Average False Rate)) by the profile updating based on threshold. At the same time, the impacts of the different training sample size, keystroke count in the sample, features, value of n of the n-graphs on the system performance were illustrated through experiments, and the reference values of these parameters were given in this paper. On the other hand, the experiment results illustrated that the numerical normalization method did none contribution to the system performance.