Trajectory analysis for user verification and recognition

For many computer activities, user verification is necessary before the system will authorize access. The objective of verification is to separate genuine account owners from intruders or miscreants. In this paper, we propose a general user verification approach based on user trajectories. A trajectory consists of a sequence of coordinated inputs. We study several kinds of trajectories, including on-line game traces, mouse traces, handwritten characters, and traces of the movements of animals in their natural environments. The proposed approach, which does not require any extra action by account users, is designed to prevent the possible copying or duplication of information by unauthorized users or automatic programs, such as bots. Specifically, the approach focuses on finding the hidden patterns embedded in the trajectories produced by account users. We utilize a Markov chain model with a Gaussian distribution in its transition to describe trajectory behavior. To distinguish between two trajectories, we introduce a novel dissimilarity measure combined with a manifold learned tuning technique to capture the pairwise relationship between the two trajectories. Based on that pairwise relationship, we plug-in effective classification or clustering methods to detect attempts to gain unauthorized access. The method can also be applied to the task of recognition, and used to predict the type of trajectory without the user's pre-defined identity. Our experiment results demonstrate that, the proposed method can perform better, or is competitive to existing state-of-the-art approaches, for both of the verification and recognition tasks.