CENTER-ASSISTED PERSONAL GAIT AUTHENTICATION USING ORIENTATION ADVERSARIAL FEATURE EXTRACTION

This work proposes a center-assisted personal authentication system using gait signals collected from inertial measurement units embedded in many wearable or mobile devices. The system utilizes a central server (with access to gait signals from multiple individuals) to learn a convolutional neural network (CNN) based feature extractor, and local wearable devices that utilize this feature extractor to perform personal authentication using a one-class classifier. In many applications (such as smartphones), the collected gait signals may depend strongly on the orientation of the devices, and the classification accuracy is often affected by the ability to remove this dependency. Hence, we propose a CNN-based feature extractor that can simultaneously maximize the ability to discriminate between different individuals' gait signals while minimizing the ability to determine the orientation associated with the signal. This is referred to as the orientation-adversarial IDNet architecture. By utilizing wild data gathered from our university campus, we show that the removal of the orientation dependency yields improved personal authentication accuracy compared to the approach adopted by the original IDNet.