Enhancing User Authentication with Wearable Data: Evaluating Sampling Strategies and Classifiers Using Apple Watch Sensor Logs

Maharage Nisansala Sevwandi Perera, Allam Shehata, Chi Xu, Xiang Li, Franziska Zimmer, Ryosuke Kobayashi, Mhd Irvan, Shigetomi Yamaguchi, Yasushi Yagi

Abstract: Wearable devices provide behavioral biometric data useful for continuous and unobtrusive user authentication. Challenges such as class imbalance and model selection can impact performance in multi-user scenarios. This study examines sampling strategies and classifiers on Apple Watch sensor data from 100 users, using a one-vs-rest framework to identify a target user among others. To address class imbalance, we evaluate the Synthetic Minority Over-sampling Technique (SMOTE) and Random Under-sampling in conjunction with Random Forest, XGBoost, and Logistic Regression classifiers. Our results show that XGBoost with weight balancing on the original dataset outperforms sampling methods, achieving 99% average accuracy and 94% F1-score with 30 users. Moreover, when we extended experiments with XGBoost to 100 users, we observed that XGBoost maintains superior performance, with 99% accuracy and 90% F1-score for 100 users. These findings underscore the significance of model selection and the handling of imbalances in developing scalable and accurate wearable-based authentication systems.