Your data in your hands: Privacy-preserving user behavior models for context computation

Modern smartphone applications rely on contextual information while providing the users with relevant and timely content and services. One way of generating such contextual information is by employing learning systems to model user behavior. Motion-based sensors, such as the accelerometer or gyroscope, have been previously employed for recognizing predefined high-level physical activities such as climbing stairs, jogging, or driving. In practice, human activities are highly diverse and unsupervised methods must be used to expose complex behavioral characteristics that are user-centric. This paper proposes a novel machine learning model for user authentication and trust that is continuously assessing the user activities in an effort to expose deviations from known training data. The goal is to export this trust score as a contextual input to mobile apps for detection of unauthorized access, fraudulent transactions, the progress of a disease, or other behavioral changes such as stage fright, intoxicated behavior, or mood changes. All collected data and generated models of the user remains on the smartphone, and only the score needs to be revealed to the apps. As a result, the user controls the data without the need to share with any remote entity. The paper presents preliminary performance results of this technique.