User Authentication and Identification on Smart Glasses with Motion Sensors

Sumeyye Agac, Özlem Durmaz Incel

Abstract: As the popularity of wearable devices increases, it is becoming crucial to provide authentication on these personal devices and prevent unauthorized access. Moreover, user identification can be important when multiple users use a device to provide a personal experience. The use of classical authentication methods, such as PINs, may not be possible, particularly on head-mounted displays or smart glasses. An alternative approach is to use behavioral biometrics using the motion sensors embedded in such devices to capture the unique movement characteristics of a user. This paper investigates the authentication and identification performance using different motion sensors alone and in combination on a smart glass. We collected a dataset from 17 participants while performing circle, up-down, tilt, triangle, turn, and square gestures. We approach both the authentication and identification problems as a classification problem and compare the performance of eight classifiers for each gesture. The results reveal that utilizing only the rotation vector and geomagnetic sensor achieves better authentication and identification performance than using only the accelerometer and gyroscope. For authentication, on average, the Adaboost algorithm with data from the rotation vector and geomagnetic sensors on the triangle gesture achieved the best equal error rate (1.3%). For identification, the best f1-score (99.3%) is obtained using the Random Forest classifier with the three different sensor combinations for up-down and square gestures.