Keystroke Biometric Authentication System Experimentation
Alpha Amatya, James Aliperti, Thomas Mariutto, Ankoor S. Shah, Michael Warren, Robert S. Zack, Charles C. Tappert · 2009
System has the ability to identify with a high degree of accuracy the typing characteristics that are unique to each individual. The system consists of three interrelated components: a java applet which collects raw keystroke data over the internet, a feature extractor, and a pattern classifier. The current study enhanced the system in several ways and developed a method of creating receiver operating characteristic (ROC) curves from the non-parametric nearest-neighbor classification procedure. The authentication system performance was increased, resulting in a decrease in the error rate by a third from 4.4 % to 2.9%. A comparison of the 1, 3, and 5-nearest-neighbor procedures indicated a significant further increase in performance in going from 1 to 3 neighbors but little change in going from 3 to 5. The ROC curves indicate potential operating points for deployed systems. Preliminary experiments were also conducted to explore the potential of increased training and “strong ” versus “weak ” enrollment.