The Keystroke Authentication System Based On Support Vector Machine

With the development of the society, the progress of science and technology has brought the benefit to people’s life, such as electronic commerce that combines the traditional commercial to the advanced network technology have completely changed the people’s way of life. People can complete the transaction at home. Online transactions bring us convenient to live, but people obviously can’t satisfy the demands for network security when we only use the account and password to complete the online trading identity recognition. Therefore people now pay more attention to the recognition technology. The reliability of traditional identity recognition technology security is weak, so the reliable biometric identification technology gradually come into our sight. Keystroke certification as a key member in the field of biometric identification, has been widely studied in recent years. The traditional keystroke authentication used the duration of the keystroke and migration time of the button to train the model. When we check the on the experimental results, this simple extraction keystroke biological features is not enough to describe the users’ behavior. In addition, it remains to be fastidious that the sample data which used to verify the effectiveness of the experimental model can reflect the user behavior in real scene. This is the reason that the field of keystroke certification was studied for a long time, but we could hardly see the commercial product.In this paper, according to the demand of the business of the company, According to the improvements on previous keystroke authentication algorithm, we complete the keystroke auxiliary certification system. The main content of the article are follow:(1) according to the standard of the software design, I completed the keystroke certification system architecture and interface design. in addition, this system complete three kinds of keystroke identification algorithm so that we can comparing the support vector machine algorithm with the other two algorithms; I also completed system design and implementation of process, the system mainly includes data collection and data preprocessing, feature extraction, training model, model prediction, model storage module. I design the data upload protocol, this test is related to large number of accounts, in order to protect the user’s privacy, I need to design a special password transfer protocol;In order to verify the improved SVM algorithm is proposed in this paper, I finish two basic keystroke model algorithm. The Bayesian classification algorithm and the Manhattan distance is used to test the performance of the improved algorithm model.(2) The basic principle of support vector machine (SVM) algorithm the derivation process is emphasized, we also focus on the basic concepts involved in the algorithm and the operation of process of algorithms. I try to avoid and control the actual engineering problems. I eliminated the former features which previous experiments used through the experiments, actually I choose the features which reflect more significant difference all depending on the result of the experiment. Support vector machine was optimized by the internal data processing so that the efficiency of the model generation and model prediction is improved. With the help of grid search method, I optimize the model parameters to achieve the automatic optimization. I also analysis the experimental results, the user sample characteristics, such as the length of the password, log in frequency, on time and the relationship between user account risk, such as the results of the password length is moderate, we conclude that the moderate length of password can effectively prevent their account from the intruder’s attack.(3)I compare the improved algorithm to two other algorithms, and present subsequent improvement ideas.In order to test the trained model in this system, I spent10days to collected the data, the user completed keystroke behavior with no perception, I think this method can effective reflect the user’s keystrokes behavior. The dataset which on the basis of the real data is divided into training set and test set, they are used to train model and test model. According to the same training set and test set, the accuracy of Manhattan distance algorithm prediction on positive samples is82.87%, the accuracy of negative samples prediction is81.67%; The accuracy of Bayesian classification algorithm prediction on positive samples is87.78%, the accuracy of negative samples prediction is86.10%; The accuracy of support vector machine algorithm prediction on positive samples is89.94%, the accuracy of negative samples prediction is95.80%. The experimental results show that the support vector machine (SVM) algorithm achieve the ideal recognition accuracy.