Isolated forest in keystroke dynamics-based authentication: Only normal instances available for training

Keystroke dynamics, which is a biometric characteristic that depends on typing style of users. In the past thirty years, dozens of classifiers have been proposed for distinguishing people using keystroke dynamics; many have obtained excellent results in evaluation. However, a more common case is that only normal instances are available and none of the rare classes are observed. It leads us to use one type model called one-class model that can only use normal instances as training sets to detect anomalies. In this paper we apply a new outlier detection algorithms, which have not been used for keystroke dynamics before: Isolation Forest (iForest). We use the existing database to test the proposed approach and it shows better performance than two other approach, especially when the number of training sample is less, both concerning accuracy and time complexity. We also research the effect of the iForest's parameters on the performance of the algorithm with small sample size. Finally, based on the original dataset we generated two new features to analyze its performance of different algorithms and we obtained nearly 0.98 Area Under Curve (AUC) of iForest.