Research on neural network algorithms for user dynamic signature recognition in the space of multidimensional curve samples, in comparison with optimal detection–discrimination algorithms for multidimensional signals
Year: 2026
Type: article
Abstract: Purpose of research . The widespread adoption of dynamic signatures in various biometric technology applicationssupported by clearly defined legal procedures in many countries-drives significant attention toward the reliability of corresponding biometric authentication algorithms. While dynamic signatures are partially free from the drawbacks inherent in static signatures, the problem of authentication reliability remains critical due to the complex interplay of heterogeneous factors. Therefore, the aim of this study is to improve the reliability of user authentication based on the dynamic signature using experimental structural and parametric synthesis of problem-oriented neural networks and comparison with classical detection-discrimination algorithms for multidimensional signals. Methods . The proposed method involves comprehensive identification of the user's dynamic signature in the sample space of multidimensional curves by means of parallel recognition of curve fragments using multiple detectors/classifiers, followed by integration and analysis of the results. Results . Neural network algorithms for identifying the user’s dynamic signature in the sample space of multidimensional curves were experimentally studied and compared with optimal detection-discrimination algorithms for multidimensional signals. The experiments demonstrated that 3–5 key parameters-including two stylus coordinates on the tablet plane, screen pressure, and stylus velocity vectors-ensure acceptable identification reliability in the range of 0,8 to 0,95 for a small number of users, and maintain a reliability level of about 0,7 with unlimited user scaling. The average gain in accuracy from using the developed models and algorithms, compared to statistical methods, amounted to 25– 35%, and compared to metric methods, 5–15%. Conclusion . To achieve the required reliability of user authentication, hardware-software identification models for dynamic signatures should be decomposed into groups with a limited number of users. There exists an optimal combination of algorithms that delivers maximum accuracy in result integration: Euclidean metric, correlation-based, and neural network classifiers.
Authors: A. K. Tantserov, E. A. Danilov
Institution: Penza State Technological University
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7126090668