An improved k-NN classification method with application to keystroke biometric authentication

The k-Nearest-Neighbor (k-NN) pattern recognition procedure is a nonparametric method of classifying patterns based on the nearest training samples in feature space. In biometric authentication systems, the Receiver Operating Characteristic (ROC) curve is a graphical representation of the trade-off between the False Accept Rate (FAR) and the False Reject Rate (FRR), and the ROC curve is important in describing the performance of the system and in determining where to set the operating point. While the derivation of ROC curves from parametric classification procedures is well understood, little is known about obtaining ROC curves from the k-NN classification procedure. This study improves the k-NN classification method by developing new techniques for obtaining ROC curves directly from k-NN results, thereby contributing to the field more flexible, inter-related two-parameter methods (k plus a threshold) rather than simply varying a single parameter k or a single ROC derivation threshold. A vector-difference model and data from a keystroke biometric authentication system are used as a vehicle to explore these techniques. An additional contribution of the study is the exploration of “strong” versus “weak” training and its effect on ROC curves derived from the k-NN method