Behavioral Biometrics for Continuous Authentication on Mobile Devices: Anomaly Detection through Keystroke Dynamics with Machine Learning
Kelvin S. Lopes, Davi O. Lopes, Wesley G. P. Pavanello, Cesar Lincoln Cavalcante Mattos, Jarelio G. da S. Filho, José D. C. Neto, Rafael L. Gomes, Nicksson C. A. de Freitas, Emanuel Bezerra Rodrigues · 2023
The popularity of mobile devices generates the need for security solutions to ensure user identity. One approach to deal with this scenario is behavioral biometric for continuous authentication (BBCA), which has become increasingly known with advances in hardware and data science technologies. However, this approach still lacks greater robustness in how to model the user’s behavioral biometrics as well as maximize the effectiveness of this authentication. It is particularly important to note in this context that this paper presents an approach to BBCA using machine learning (ML) techniques integrated with sliding windows. Accordingly, biometric data collected from keystroke dynamics typing activities on mobile devices were used to identify genuine users, random (unskilled) imposters and skilled imposters. The ML models for impostor identification followed an anomaly detection (AD) approach, where only genuine data is available at training time. In addition, the use of sliding windows allowed the inclusion of the temporal dimension of the task. The results obtained indicate that the proposed solution has a practical feasibility in terms of its suitability to perform user identification, specifically, the KNN model stood out by achieving a superior performance in the value window 4. In this configuration, it achieved a score of 94% for the F1-score metric in the random imposter scenario, and 93% in the skilled imposter scenario.