Robust Keystroke Behavior Features for Continuous User Authentication for Online Fraud Detection

Abstract

Recently, behavioral biometric-based user authentication methods, such as keystroke dynamics, have become a popular alternative to improve security of online platforms, due to their non-invasive nature. However, currently there are very few behavioral biometric authentication methods that provide non-invasive continuous user authentication for online education platforms, resulting in frequent network intrusion and online assessment fraud. Existing approaches mostly analyze the typing behavior of users using a fixed sequence of characters. Furthermore, a better set of features are required to reduce false positive rate for satisfactory performance to prevent online fraud. Existing behavioral analysis methods also mostly rely on conventional machine learning approaches despite recent advancement in deep learning approaches. We identify a set of keystroke behavioral biometric features that yield satisfactory performance by identifying most frequently used features. We also collect new free-form keystroke behavior data during online assessment activities and develop non-invasive continuous authentication methods for free-form text behavior analysis using deep learning approaches. We also compare performance between deep learning and conventional machine learning approaches and evaluate the robustness of the most frequently used features. Result analysis shows that deep learning approaches outperform machine learning approaches on most frequently used feature set. Furthermore, it is found that the identified feature set is robust and results in satisfactory performance in deep learning approaches.