Continuous authentication through gait analysis on a wrist-worn device

Being distinctive of every individual, gait can be used as a biometric feature to authenticate the owner of a wearable device. This paper proposes and evaluates an authentication method that relies on the acceleration signal acquired at the user's wrist. During the training phase, the wrist-worn device automatically learns the gait patterns of the legitimate user, by exploiting a set of acceleration-based indicators. Subsequently, unauthorized users are detected by observing the occurrence of anomalous gait patterns. Experimental results carried out with 20 volunteers show that the proposed method is able to recognize the legitimate user with an equal error rate of ∼2.5%. The method is sufficiently lightweight to be executed in real time on a wearable device with limited resources. This enables continuous authentication without requiring the presence of an external device (e.g., a smartphone). Furthermore, the provided evaluation of power consumption shows that the completely on-node solution is also more energy efficient with respect to off-loading computation to an external device.