Comparing classification algorithm for mouse dynamics based user identification

Mouse dynamics is the process of identifying individual users based on their mouse operating behaviors. Many classification algorithms have been proposed for checking users' identity, thus it is natural to ask how well each classifier performs and how various classifiers compare to each other (e.g., to identify promising research directions). Unfortunately, we cannot conduct a valid comparison of classifiers based on the results in the literature because of inconsistent datasets, experimental conditions and methodologies across studies. The objective of this study is to obtain a dataset, to develop an objective and repeatable evaluation procedure, and to measure the performance of a range of classifiers on an equal basis. Using data from 26 subjects conducting a fixed mouse-operation task 300 times each, we implemented and evaluated twelve classifiers from the mouse-dynamics and pattern-recognition literature for a user identification task. The four top-performing classifiers achieve false-negative identification rates between 17.31% and 23.72%. The results, along with the evaluation methodology, constitute a benchmark for comparing classifiers and measuring progress for the user identification problem.