Manifold learning for user profiling and identity verification using motion sensors

Abstract Mobile devices are becoming ubiquitous and being increasingly used for data-sensitive activities such as communication, personal media storage, and banking. The protection of such data commonly relies on passwords and biometric traits such as fingerprints. These methods perform the user authentication sporadically and often require action from the user, which may make them susceptible to spoofing attacks. This scenario can be mitigated if we bring to bear motion-sensing based methods for authentication, which operate continuously and without requiring user action, hence are harder to attack. Such methods could be used allied with traditional authentication methods or on their own. This paper explores this idea in a novel user-agnostic approach for identity verification based on motion traits acquired by mobile sensors. The proposed approach does not require user-specific training before deployment in mobile devices nor does it require any extra sensor in the device. This solution is capable of learning a user profiling manifold from a small user subset and extend it to unknown users. We validated the proposal on two public datasets. The reported experiments demonstrate remarkable results under a cross-dataset protocol and an open-set setup. Moreover, we performed several analyses aiming at answering critical questions of a biometric method and the presented solution.