Continuous Typist Verification using Machine Learning

Abstract

A keyboard is a simple input device. Its function is to send keystroke information to the computer (or other device) to which it is attached. Normally this information is employed solely to produce text, but it can also be utilized as part of an authentication system. Typist verification exploits a typist's patterns to check whether they are who they say they are, even after standard authentication schemes have confirmed their identity. This thesis investigates whether typists behave in a sufficiently unique yet consistent manner to enable an effective level of verification based on their typing patterns.

Typist verification depends on more than the typist's behaviour. The quality of the patterns and the algorithms used to compare them also determine how accurately verification is performed. This thesis sheds light on all technical aspects of the problem, including data collection, feature identification and extraction, and sample classification.

A dataset has been collected that is comparable in size, timing accuracy and content to others in the field, with one important exception: it is derived from real emails, rather than samples collected in an artificial setting. This dataset is used to gain insight into what features distinguish typists from one another. The features and dataset are used to train learning algorithms that make judgements on the origin of previously unseen typing samples. These algorithms use 'one-class classification'; they make predictions for a particular user having been trained on only that user's patterns.

This thesis examines many one-class classification algorithms, including ones designed specifically for typist verification. New algorithms and features are proposed to increase speed and accuracy. The best method proposed performs at the state of the art in terms of classification accuracy, while decreasing the time taken for a prediction from minutes to seconds, and - more importantly - without requiring any negative data from other users. Also, it is general: it applies not only to typist verification, but to any other one-class classification problem.

Overall, this thesis concludes that typist verification can be considered a useful biometric technique.