| Abstract [eng] |
Biometric authentication has emerged as a crucial technology for enhancing digital security, with keystroke dynamics providing a cost-effective and accessible behavioral biometric solution. This thesis evaluates and compares various machine learning approaches for keystroke dynamics-based user authentication, and proposes solutions to improve the accuracy of existing authentication techniques. Using datasets such as CMU and KeyRecs, various classifiers, including decision trees, random forests, k-nearest neighbors, support vector machines, gradient boosting, XGBoost, and convolutional neural networks, are implemented. In addition, user classification based on time series images is introduced. The transformation of time series data into gramian angular summation field and gramian angular difference field images further enhances the analysis. The results demonstrated significant variations in classifier performance between datasets. For the CMU dataset, the voting ensemble model achieves the highest accuracy of 96.48%, whereas for the KeyRecs dataset, the convolutional neural network achieves up to 90.45% accuracy with an equal error rate as low as 0.1006. The obtained results indicate that it is meaningful to analyze keystroke dynamics on both numeric features and time series images. |
