Benchmarking keystroke authentication algorithms

Free-text keystroke dynamics is a behavioral biometric that has the strong potential to offer unobtrusive and continuous user authentication. Such behavioral biometrics are important as they may serve as an additional layer of protection over other one-stop authentication methods such as the user ID and passwords. Unfortunately, evaluation and comparison of keystroke dynamics algorithms are still lacking due to the absence of large, shared free-text datasets. In this research, we present a novel keystroke dynamics algorithm, based on kernel density estimation (KDE), and contrast it with two other state-of-the-art algorithms, namely Gunetti & Picardi's and Buffalo's SVM algorithms, using three published datasets, as well as our own new, unconstrained dataset that is an order of magnitude larger than the previous ones. We modify the algorithms when necessary such that they have comparable settings, including profile and test sample sizes. Both Gunetti & Picardi's and our own KDE algorithms have performed much better than Buffalo's SVM algorithm. Although much simpler, the newly developed KDE algorithm is shown to perform similarly as Gunetti & Picardi's algorithm on the three constrained datasets, but the best on our new unconstrained dataset. All three algorithms perform significantly better on the three prior datasets, which are constrained in one way or another, than our new dataset, which is truly unconstrained. This highlights the importance of our unconstrained dataset in representing the real-world scenarios for keystroke dynamics. Lastly, the new KDE algorithm degrades the least in performance on our new dataset.