Keystroke dynamics recognition using cluster-based prediction ellipses

Abstract

Keystroke dynamics recognition is a biometric approach that enables user identification based on individual typing patterns. The relevance of the study is determined by the need to improve the accuracy of keystroke dynamics recognition under conditions of variability, heterogeneity, and deviation of data from a multivariate normal distribution. The object of this research is the keystroke dynamics recognition process, while the subject of the research is a mathematical model based on cluster-based prediction ellipses constructed in a two-variate feature space. The purpose of the article is to improve the accuracy of keystroke dynamics recognition by applying cluster-based prediction ellipses that take local data structures into account. The objectives of the study are to form cluster-based prediction ellipses constructed on local training samples and to compare the proposed approach with other one-class recognition methods. The methods of the study include Mardia’s test for assessing multivariate normality, the Box-Cox transformation for data normalization, k-means clustering, the elbow method and cluster size analysis, the construction of prediction ellipses based on the squared Mahalanobis distance, and robust covariance estimation based on the minimum covariance determinant. The scientific novelty consists in developing a cluster-based approach to constructing local prediction ellipses for keystroke dynamics recognition, which makes it possible to account for data heterogeneity and form more precise decision boundaries. The practical significance of the study lies in the possibility of improving the reliability of biometric user authentication systems without switching to more complex multiclass models. The results of the study demonstrate that the proposed approach significantly outperforms the one-class support vector machine and the prediction ellipse constructed for non-Gaussian data, while also providing better recognition of samples belonging to another class and recognition accuracy comparable to the prediction ellipse for normalized data. This improvement is achieved through adaptive modeling of local data distributions and more precise decision boundaries in the feature space. The conclusions confirm the effectiveness of clustering as a preprocessing stage for keystroke dynamics recognition based on cluster-based prediction ellipses and demonstrate the suitability of adaptive local modeling for heterogeneous keystroke dynamics data. Further research may focus on the application of alternative clustering methods, as well as on the use of advanced data normalization methods, such as the Johnson transformation. In addition, extending the feature space to a higher dimensionality by incorporating a larger number of keystroke dynamics characteristics may further improve recognition accuracy.