Behavioral Authentication Using Activity-Based Model Segregation in Convolutional Neural Networks for Mobile Sensor Data
Thapanapong Rukkanchanunt, Saksinkarn Petchkuliinda, Hideya Ochiai · 2024
This study explores user authentication through Convolutional Neural Networks (CNN) on mobile sensor data, aiming to enhance security in mobile computing. We utilized the MobiAct dataset to authenticate individuals based on behav-ioral patterns captured by smartphone sensors. Our approach compares different data segmentation strategies, focusing on the impact of window size and sliding step on four performance metrics: accuracy, precision, recall, and Fl-score. Results demonstrate that larger window sizes generally lead to increased accuracy. We examine the implications of activity-based model segregation, leading to improvements in precision and overall model stability. We also confirm that walking is most suitable activity in identifying individual. This research contributes to the growing field of behavioral-based user authentication, showcasing the potential of using inherent motion sensor data in smartphones for creating more personalized and secure user experiences.