Stability in Anomaly Detection for Keystroke Dynamics: Exploring the Possibility with Plateau Regression of an Individual's Keystroke Dynamics Changing Over Time
Presentation Type
Poster
Student
Yes
Track
Other
Abstract
Anomaly detection methods of keystroke dynamics have been proposed for enhanced biometric security for passwords. This study was proposed to confirm the underlying assumptions of these anomaly detectors and see if an individual's keystroke dynamics changing over time, which would skew the training of anomaly detection algorithms and lead to low adequacy. The data used for this study was from a public benchmark data set from Killouhy and Maxion. [5] The data consisted of 51 subjects from the campus of Carnegie Mellon, that completed 8 data-collection sessions completing 50 password repetitions that had to be the password typed correctly. The data was analyzed on the session level for the total time to type the password, as it best describes the entire keystroke dynamics for each individual. Summary statistics had to be derived for each session and median regression by quantile regression was implemented across session using the conditional median across the session or a linear plateau method to derive where subjects became stationary in sessions. The summary statistic was then used in a linear mixed-effect model to test for significance in the slope of the session. It was found that session has a significant decrease in slope and thus there is evidence to show an individual’s typing dynamics change over time. Recommendations on how to proceed with the anomaly detector are to explore the idea of implementing adaptive training techniques for the anomaly detector.
Start Date
2-5-2019 12:00 PM
End Date
2-5-2019 1:00 PM
Stability in Anomaly Detection for Keystroke Dynamics: Exploring the Possibility with Plateau Regression of an Individual's Keystroke Dynamics Changing Over Time
Volstorff A
Anomaly detection methods of keystroke dynamics have been proposed for enhanced biometric security for passwords. This study was proposed to confirm the underlying assumptions of these anomaly detectors and see if an individual's keystroke dynamics changing over time, which would skew the training of anomaly detection algorithms and lead to low adequacy. The data used for this study was from a public benchmark data set from Killouhy and Maxion. [5] The data consisted of 51 subjects from the campus of Carnegie Mellon, that completed 8 data-collection sessions completing 50 password repetitions that had to be the password typed correctly. The data was analyzed on the session level for the total time to type the password, as it best describes the entire keystroke dynamics for each individual. Summary statistics had to be derived for each session and median regression by quantile regression was implemented across session using the conditional median across the session or a linear plateau method to derive where subjects became stationary in sessions. The summary statistic was then used in a linear mixed-effect model to test for significance in the slope of the session. It was found that session has a significant decrease in slope and thus there is evidence to show an individual’s typing dynamics change over time. Recommendations on how to proceed with the anomaly detector are to explore the idea of implementing adaptive training techniques for the anomaly detector.