A second look at the performance of neural networks for keystroke dynamics using a publicly available dataset

Keystroke Dynamics, which is a biometric characteristic that depends on typing style of users, could be a viable alternative or a complementary technique for user authentication if tolerable error rates are achieved. Most of the earlier studies on Keystroke Dynamics were conducted with irreproducible evaluation conditions therefore comparing their experimental results are difficult, if not impossible. One of the few exceptions is the work done by Killourhy and Maxion, which made a dataset publicly available, developed a repeatable evaluation procedure and evaluated the performance of different methods using the same methodology. In their study, the error rate of neural networks was found to be one of the worst-performing. In this study, we have a second look at the performance of neural networks using the evaluation procedure and dataset same as in Killourhy and Maxion's work. We find that performance of artificial neural networks can outperform all other methods by using negative examples. We conduct comparative tests of different algorithms for training neural networks and achieve an equal error rate of 7.73% with Levenberg-Marquardt backpropagation network, which is better than equal error rate of the best-performing method in Killourhy and Maxion's work.