Towards Passive Authentication using Inertia Variations: An Experimental Study on Smartphones

Abstract

Passive biometrics and behavioral analytics seek to identify users based on their unique patterns of activities. In this paper, we test the feasibility of using time-varying inertia data as passive biometrics to be used for user identification and authentication. We present a deep learning model for inertia pattern recognition that achieved a high accuracy of 87.17%. A fully-connected sequential deep neural network was trained on 6730 sensor data samples, each having 15 features: triaxial measurements from accelerometer, gyroscope, magnetometer, and rotational vector. We further discuss the potential impact of inertia pattern recognition for user identification and authentication.