Many of our activities on computer need a verification step for authorized access. The goal of verification is to tell apart the true account owner from intruders. We propose a general approach for user verification based on user trajectory inputs. The approach is labor-free for users and is likely to avoid the possible copy or simulation from other non-authorized users or even automatic programs like bots. Our study focuses on finding the hidden pattern on the trajectories produced by account users. We employ a Markov chain model with Gaussian distribution in its transitions to describe the behavior on the trajectory. To distinguish between two trajectories, we propose a novel dissimilarity measure combined with a manifold learnt tuning for catching the pairwise relationship. Based on the pairwise relationship, we plug-in any effective classification or clustering methods for the detection of unauthorized access. The result shows that the proposed method can accurately verify the user identity. The similar method may also be applied for the task of recognition, predicting the trajectory type without pre-defined identity.

Hsing-Kuo Pao, Hong-Yi Lin, and Kuan-Ta Chen, "Trajectory based Behavior Analysis for User Verification," In Proceedings of IDEAL 2010, Sep 2010.

Full Text (PDF Format, 335 KB)

Slides (PDF Format, 3063 KB)