Open Access

ARTICLE

Mobile Touch Dynamics–Based User Classification Using Machine Learning and Fusion Techniques

Animaw Kerie Aseres^1,2,*, Asrat Mulatu Beyene^3,2, Lemlem Kassa Tegegne^1,2

1 Department of Software Engineering, Addis Ababa Science and Technology University, Addis Ababa, Ethiopia
2 High Performance Computing and Big Data Analytics Center of Excellence, Addis Ababa Science and Technology University, Addis Ababa, Ethiopia
3 Department of Electrical and Computer Engineering, Addis Ababa Science and Technology University, Addis Ababa, Ethiopia

* Corresponding Author: Animaw Kerie Aseres. Email:

Journal of Cyber Security 2026, 8, 559-576. https://doi.org/10.32604/jcs.2026.086559

Received 02 June 2026; Accepted 23 July 2026; Issue published 21 August 2026

Abstract

Conventional multi-factor and one-time authentication approaches, such as passwords and one-time passwords (OTPs), have become increasingly vulnerable to advanced attack methods, motivating the need for continuous authentication (CA) systems that can verify user identity throughout an active session rather than only at login. For such a system to be effective, it must analyze user behavior reliably and in real time. This paper presents a novel approach to implementing CA on mobile devices using tap and swipe behavioral biometrics combined with machine learning (ML) and multimodal fusion. The dataset was collected from 400 volunteer participants using BahriApp, a custom-developed Android application (Bahri meaning “behavior” in Amharic). Several models were then trained to classify user identity, including Logistic Regression, SVM, Random Forest, and deep learning models (1D-CNN and LSTM), evaluated with feature-level and decision-level fusion of the tap and swipe datasets. Feature-level fusion using LSTM performed best, achieving 97.1% accuracy and a 4.4% Equal Error Rate (EER), balancing usability and security; weighted decision-level fusion achieved 96.5% accuracy and a 4.8% EER. The main contributions of this paper are: a custom behavioral biometrics dataset constructed from African participants; comparative benchmarks across machine learning models and fusion strategies; and insights that inform the design of an adaptive continuous authentication scheme. The study’s main limitation is dataset size and session coverage: of the 400 participants, only 29.75% and 20.75% provided 15 or more sessions in the tap and swipe datasets, respectively. Future work will focus on larger and more diverse multimodal datasets.

---

Keywords

---