[Paper Review] Evaluation of a User Authentication Schema Using Behavioral Biometrics and Machine Learning

Laura Pryor, Jacob Mallet+4 more|arXiv (Cornell University)|May 7, 2022

User Authentication and Security Systems4 citations

TL;DR

This paper proposes a multi-modal user authentication system combining touch dynamics and phone movement behaviors using machine learning for enhanced security on mobile devices. Evaluated on two public datasets using seven algorithms, the scheme achieved up to 86% accuracy, demonstrating the effectiveness of behavioral biometrics as a robust alternative to traditional password- or physiology-based methods.

ABSTRACT

The amount of secure data being stored on mobile devices has grown immensely in recent years. However, the security measures protecting this data have stayed static, with few improvements being done to the vulnerabilities of current authentication methods such as physiological biometrics or passwords. Instead of these methods, behavioral biometrics has recently been researched as a solution to these vulnerable authentication methods. In this study, we aim to contribute to the research being done on behavioral biometrics by creating and evaluating a user authentication scheme using behavioral biometrics. The behavioral biometrics used in this study include touch dynamics and phone movement, and we evaluate the performance of different single-modal and multi-modal combinations of the two biometrics. Using two publicly available datasets - BioIdent and Hand Movement Orientation and Grasp (H-MOG), this study uses seven common machine learning algorithms to evaluate performance. The algorithms used in the evaluation include Random Forest, Support Vector Machine, K-Nearest Neighbor, Naive Bayes, Logistic Regression, Multilayer Perceptron, and Long Short-Term Memory Recurrent Neural Networks, with accuracy rates reaching as high as 86%.

Motivation & Objective

• To address the limitations of traditional mobile authentication methods such as passwords and physiological biometrics, which are vulnerable to spoofing and static security measures.
• To explore the feasibility and performance of behavioral biometrics—specifically touch dynamics and phone movement—as a continuous and adaptive authentication mechanism.
• To evaluate single-modal and multi-modal combinations of touch and motion biometrics using diverse machine learning algorithms.
• To contribute empirical evidence on the effectiveness of behavioral biometrics in real-world mobile authentication scenarios.

Proposed method

• The study uses two publicly available datasets: BioIdent for touch dynamics and H-MOG for phone movement and orientation.
• Feature extraction is performed on touch interactions (e.g., tap duration, pressure, swipe speed) and motion patterns (e.g., acceleration, angular velocity) from smartphone sensors.
• Seven machine learning models are trained and evaluated: Random Forest, Support Vector Machine, K-Nearest Neighbor, Naive Bayes, Logistic Regression, Multilayer Perceptron, and Long Short-Term Memory (LSTM) networks.
• Performance is assessed using classification accuracy, with cross-validation to ensure robustness across subjects and sessions.
• Multi-modal fusion is implemented by concatenating features from both touch and motion modalities before feeding them into the classifiers.
• The evaluation includes both single-modal (touch-only or motion-only) and multi-modal (touch + motion) configurations to compare performance.

Experimental results

Research questions

• RQ1 How does the performance of touch dynamics alone compare to phone movement alone in user authentication using machine learning?
• RQ2 What is the impact of combining touch and motion biometrics on authentication accuracy compared to single-modal approaches?
• RQ3 Which machine learning algorithm achieves the highest accuracy in classifying user identity based on behavioral biometric features?
• RQ4 How do different feature combinations from touch and motion data affect model generalization and robustness across subjects?
• RQ5 Can behavioral biometrics serve as a viable, continuous, and secure alternative to traditional password- or physiology-based authentication?

Key findings

• The multi-modal approach combining touch dynamics and phone movement achieved the highest authentication accuracy of 86% across the evaluated datasets.
• Among the seven machine learning models, Long Short-Term Memory (LSTM) networks demonstrated strong performance, particularly in capturing temporal patterns in behavioral sequences.
• Random Forest and Support Vector Machine also showed high accuracy, indicating strong generalization across diverse user behaviors.
• Single-modal systems performed adequately but consistently underperformed compared to the multi-modal fusion approach.
• The results confirm that behavioral biometrics, when combined with machine learning, can provide a reliable and adaptive authentication mechanism for mobile devices.