Personal identification by Keystroke Pattern for login security

This thesis discusses the Neural Network (NN) approach in identifying personnel through keystroke behavior in the login session. The keystroke rhythm that falls in the behavioral biometric has a unique pattern for each individual. Therefore, these heterogeneous data obtained from normal behavior users can be used to detect intruders in a computer system. The keystroke behavior was captured in the form of time within the duration between the pressing and releasing of key was recorded during the login session. Ten frequent loggers were chosen for the experiments. The data obtained were presented to NN for pattern learning and classifying the strings of characters. The backpropagation (BP) model was implemented to identify the keystroke patterns for each class.Various architectures were employed in the SP training to achieve the best recognition rate. Several features that influence the network were considered. The experiment involved the slicing of input data and the determination of the number of hidden units. Several other factors such as momentum, learning rate and various weight initialization were used for comparison. Three types of weight initialization were used, including Nguyen-Widrow (NW), Random and Genetic Algorithm (GA). The experiment showed that the recognition of 97% was achieved using NW weight initialization with 10 hidden units. Further experiments with Improved Error Function (IEF) in standard SP has showed better results with 100% recognition on both train and test data set compared to previous experiment. The results of this study were compared with Chambers's (1990) and Obaidat's (1994) work. Chambers used the data set similar to the data used in this experiment and obtained 90.5% recognition through Inductive Learning Classifier method, while Obaidat used standard BP with 6 classes and obtained 97.5% recognition.

Text FSKTM 2001 1 IR.pdf Download (1MB)

Actions (login required)

View Item