Zhihua Ding

With the rapid growth of wearable devices and the Internet of Things (IoT), users increasingly rely on smartwatches and fitness trackers for services in connected environments. These devices collect sensitive physiological and behavioral data, requiring reliable user authentication to ensure secure access by authorized users. However, few-shot scenarios with significant intraclass variability and limited training data make it challenging for models to capture complex feature interactions. In this article, we propose a ratio-enhanced feature interaction (RE-FI) framework designed for wearable running data. RE-FI systematically constructs pairwise product and ratio-based cross features to explicitly model second-order interactions and applies information gain (IG)-based feature selection to retain the most discriminative features. We provide theoretical justification for these design choices across linear models, tree-based classifiers, and deep neural networks (DNNs). We validate RE-FI on real-world running data using nine representative classifiers. Results show that RE-FI consistently outperforms the baseline feature set across all classifiers, achieving an average accuracy improvement of 4.2% and up to 6.2% in $F1$ -score. In extreme few-shot settings with only five samples per user, RE-FI maintains robust performance, with random forest (RF) achieving an average accuracy of 86%. Overall, this work offers a computationally simple feature interaction solution that is compatible with diverse supervised models on session-level wearable statistical features and provides a principled approach for wearable security and personalized health monitoring in IoT applications.