One-class SVM for biometric authentication by keystroke dynamics for remote evaluation

Remote skills assessment in distance education needs individual identification in distinguishing between candidates and impostors. Keystroke dynamics is a behavioral biometrics which can be used to identify them. To expect lower error rate, behaviors should be as natural and consistent as possible. The unique identifier assigned to students at their registration seems appropriate but the classification method applied for this case of anomaly detection must be robust even with a lower signature number. In this paper, we first explain how we construct our own dataset. Three methods of selecting Gaussian kernel parameters for one-class support vector machine are subsequently studied regarding the targeted application constraints. The results show that an indirect method as distance to farthest neighbor cannot be used because some signature features have multimodal and dispersed distributions. A method is then proposed based on the selection of the parameters via detecting the "tightness" of the decision boundaries and uses a greedy search. Its performances are compared to those of a grid search method using LibSVM. The results show that the proposed method is more robust when the signatures number decreases and better and more stable in detecting impostors.