Account-Sharing Detection Through Keystroke Dynamics Analysis

Abstract

Account sharing refers to a situation where multiple individuals share a Web site account to avoid paying a fee or providing personal information. As a result of account sharing, service providers lose revenue, underestimate membership, and have impaired understanding of their customers. A generic framework for detecting account sharing is proposed, using keystroke dynamics. Starting with the observation that a user''''''''s keystroke patterns are consistent and distinct from those of other individuals, it is assumed that each user''''''''s keystroke patterns form a "cluster" in Euclidean space. The number of sharers can be estimated by the number of clusters. In this paper, the "optimal" number of clusters is estimated based on the Bayesian model-selection framework with Gaussian mixture models obtained using the variational Bayesian approach. In a case study involving 25 passwords and 16 users, the proposed approach performed well in "sharing detection," with a 2 percent false alarm rate, a 2 percent miss rate, and a "total user estimation" error of 7 percent. The proposed approach is viable and merits further investigation.