An Improved User Identification based on Keystroke-Dynamics and Transfer Learning

Abstract

Transfer learning is one of the most appreciated classification techniques, when a small size of training dataset and limited computer capabilities exist along with the presence of noisy data. Such a scenario is encountered in keystroke dynamics whereby the typing gesticulation composed to form ‘keystroke dynamics’, is used to distinguish a human. Keystroke dynamics is persistently gaining reputation as one of the most promising and cost-effective behavioral biometrics. AlexNet and ResNet are two separate types of pre-trained convolutional neural network models, normally used to apply deep transfer learning concepts within image-based systems. Conversion of keystroke data into image data and then the formation of the artificial image data by extending available data are crucial attraction of this paper. In this paper, the pre-trained models are utilized by applying fine-tuning (of parameters in some layers) and used as an end-to-end learning and classification system by us. On the keystroke dataset, the feature extraction method (support vector machine for classification) is also applied here with pre-trained models. A relative analysis of both approaches is provided and lastly, the better one is employed in the proposed (recognition based) system. Finally, 98.57% accuracy of successful recognition is recorded.