Abstract:
"
Smartwatches have become a key component within the realm of personal computing. These devices have been tightly woven with the common smartphone and the user itself. However, smartwatches lack the ability to employ well-known biometric authentication techniques such as fingerprint scanning and facial recognition due to their physical space constraints and other challenges. Gait analysis is the study of walking patterns, and it has known to be a feasible authentication technique, especially using the accelerometer and gyroscope sensors within smartphones. This study explores the feasibility of a gait-based authentication scheme that uses accelerometer and gyroscope data from a commercially available smartwatch device. A novel approach of using lightweight Siamese long-short term memory (LSTM) networks to identify unique features from two sets of sensor data is proposed for performing gait-based authentication. Furthermore, the computational costs of executing this type of authentication scheme are also explored with the proposed LSTM model architecture.
The author concludes that models trained in the aforementioned configuration would be feasible given the right amount of training data and necessary model architecture that can extract discriminative features within the sensor signals. Within this research the author evaluates 4 different Long Short-Term Memory (LSTM) based pairwise learning models with the best configuration achieving an accuracy of 94% and an Equal Error Rate of 0.26%. This configuration uses a Contrastive loss function and Cosine similarity as a distance metric. The author also assesses system performance when performing inference to this model locally, at its peak during inference the CPU usage of a Samsung Galaxy Watch 4 peaked at 51% and memory usage at 148MB."
