Feature Distribution-Based Touch Biometrics Using CNN and Siamese Networks

Ensuring secure and user-friendly authentication is important as mobile devices increasingly handle sensitive data. Traditional methods like PINs, fingerprints, and facial recognition have privacy limitations, whereas behavioral biometrics offer implicit, continuous authentication. This study presents a touch-based authentication framework leveraging feature distribution modeling with a Convolutional Neural Network (CNN)-based Siamese network. Using Kullback-Leibler (KL) divergence, we compare touch dynamics distributions across sessions to differentiate users. To address behavioral variability, we employ adaptive bandwidth tuning in kernel density estimation (KDE) for improved probability modeling. The CNN extracts embeddings from these feature distributions, while the Siamese network assesses session similarities. Unlike traditional handcrafted approaches using summary statistics, our method preserves the full statistical structure of touch interactions, improving authentication accuracy. Experimental results demonstrate competitive Equal Error Rates (EER), underscoring the potential of distribution-driven touch biometrics for mobile authentication.