What question did this study set out to answer?

The aim is to evaluate and compare the effectiveness of different keystroke-dynamics authentication methods using machine learning and deep learning approaches.

May 17, 2026Black Sea Journal of Engineering and Science0 citationsOpen Access

Keystroke Dynamics for Authentication: A Systematic Comparative Analysis of Machine Learning, Deep Learning, Hybrid, and Quantum Approaches

NGNebil Vural GündoğanIşık University BÇBarış ÇeliktaşMEF University

Key Points

The aim is to evaluate and compare the effectiveness of different keystroke-dynamics authentication methods using machine learning and deep learning approaches.
Conducted a systematic review of existing literature on keystroke dynamics authentication across various device contexts.
Analyzed common datasets and evaluated models using metrics like EER, FAR, and FRR.
Assessed risks and proposed mitigations based on findings from the review.
Modern deep learning models achieve an EER of approximately 0.01–2% on large, diverse datasets, outperforming classical machine learning.
Hybrid approaches that combine deep learning with classical classifiers improve user experience and robustness.
Quantum machine learning claims of near-perfect accuracy are limited to small or simulated studies and lack generalizability.

Abstract

We present a systematic review of keystroke-dynamics authentication across classical machine learning (ML), deep learning (DL), hybrid, and emerging quantum-ML approaches. The review synthesizes evidence from desktop, mobile, and web settings; compares common datasets (e.g., CMU, GREYC, Aalto, Clarkson); and evaluates models using EER, FAR, and FRR. Three findings stand out. First, dataset breadth, device context, and evaluation protocol explain much of the variance reported in the literature; on free-text inputs, modern DL (e.g., Transformers, Siamese networks) typically achieves EER ≈0.01–2% on large, diverse sets, outperforming classical ML. Second, hybrid designs, feature extraction with DL plus classical classifiers or multimodal fusion (keystroke + touch/mouse), improve robustness and user experience relative to single-modality systems. Third, claims of near-perfect accuracy from quantum-ML are confined to small or simulated studies and are not yet generalized. We map model risks to the OWASP Authentication Cheat Sheet (replay/spoofing, template security, adversarial examples, lockout usability) and outline mitigations (MFA, liveness, cancelable templates, throttling). Finally, we chart a practical pipeline and highlight near-term directions: federated learning and edge deployment for privacy/latency, explainable AI for auditability, and standardized benchmarks for fair comparison.

问 AI

Bookmark

View Full Paper