User recognition via keystroke latencies using SOM and Backpropagation Neural Network

Keystroke biometric has been widely investigated to enable an ordinary keyboard to be an authenticable input device. Most of the existing studies have analysed keystroke latencies in some frequently-used, short, fixed-length strings, e.g. username and password, to strengthen authentication in general login processes. Meanwhile, a number of studies have been recently proposed to use keystroke latency in free text domain in which a sequence of keystrokes is generally longer than username and password. However, the methods used in both short text and long text are different. General supervised classifiers which are used with a fixed number of attributes cannot be applied to the variable-length free text. In this paper, we thus propose a new method which can employ backpropagation neural network to classify the users using keystroke latency times from non-fixed length text. Our method first clusters the latency between a pair of characters or digraph, and then discovers the group of digraphs which identifies each user. Next, the cluster membership is used as input features of the networks. The results show that our method can recognize unseen users with a higher accuracy than the other methods run in our experiments.