The Partially Observable Hidden Markov Model and its Application to\n Keystroke Dynamics

The partially observable hidden Markov model is an extension of the hidden\nMarkov Model in which the hidden state is conditioned on an independent Markov\nchain. This structure is motivated by the presence of discrete metadata, such\nas an event type, that may partially reveal the hidden state but itself\nemanates from a separate process. Such a scenario is encountered in keystroke\ndynamics whereby a user's typing behavior is dependent on the text that is\ntyped. Under the assumption that the user can be in either an active or passive\nstate of typing, the keyboard key names are event types that partially reveal\nthe hidden state due to the presence of relatively longer time intervals\nbetween words and sentences than between letters of a word. Using five public\ndatasets, the proposed model is shown to consistently outperform other anomaly\ndetectors, including the standard HMM, in biometric identification and\nverification tasks and is generally preferred over the HMM in a Monte Carlo\ngoodness of fit test.\n