Comparative Study of Various Features-Mining-Based Classifiers in Different Keystroke Dynamics Datasets

Abstract

Habitual typing rhythm or keystroke dynamics is a behavioural biometric characteristic in Biometric Science relates the issue of human identification/ authentication. In 30 years of on-going research, many keystroke dynamics databases have been created on various pattern of strings (“greyc laboratory”, “.tie5Roanl”, “the brown fox”, …) taking various combination of keystroke features (flight time, dwell time) and many features-mining classification algorithms have been proposed. Many have obtained impressive results. But in evaluation process, a classifier’s average Equal Error Rates (EERs) are widely varied from 0 to 37 % on different datasets ignoring typographical errors. The question may arise, which classifier is best on which pattern of keystroke databases? To get the answer, we have started our experiment and created our own five rhythmic keystroke databases on different daily used common pattern of strings (“kolkata123”, “facebook”, “gmail.com”, “yahoo.com”, “123456”) and executed various classification algorithms in R statistical programming language, so, we can compare the performance of all the classification algorithms soundly on different datasets. We have executed 22 different classification algorithms on collected data considering various keystroke features separately. In the observation, obtained best average EER of the classifier Lorentzian is 1.86 %, where 2.33 % for Outlier Count, 3.69 % for Canberra, 5.3 % for Naïve Baysian and 8.87 % for Scaled Manhattan by taking all five patterns of strings and all combination of features in the consideration. So the adaptation of keystroke dynamics technique in any existing system increases the security level up to 98.14 %.