Extended sammon projection and wavelet kernel extreme learning machine for gait-based legitimate user identification

Smartphones have pervasively integrated into our home and work environments managing confidential information but their owners still rely on as explicit as inefficient and insecure identification processes. Therefore, if a device is stolen, a thief can have access to the owner's personal information and services though the stored password/s. To avoid such situations, this work demonstrates the possibilities of legitimate user identification in a semi-controlled environment through the built-in smartphone motion dynamics captured by two different sensors. This is a two step process: sub-activity recognition followed by user/impostor identification. Prior to the identification, Extended Sammon Projection (ESP) method is used to reduce the redundancy among the features. To validate the proposed system, we first collected data from four users walking with their device freely placed in one of their pants pockets. Through extensive experimentation, we demonstrated that time and frequency domain features, optimized by ESP to train the wavelet kernel based extreme learning machine classifier, implement an effective system to identify the legitimate user or an impostor with 97% accuracy.