Structural synthesis of neural network models for biometric recognition in the feature space of multidimensional dynamic curves

Purpose of research . The study aims to improve the reliability of biometric user identification based on dynamic signatures by developing neural network models operating in the feature space of multidimensional dynamic curves. The focus is on the structural and parametric synthesis of a classification neural network architecture using statistical, harmonic, and wavelet-transformed features extracted from the dynamic signature. Methods . The proposed identification model performs parallel recognition of multidimensional curve fragments using various methods, including statistical, metric, and neural classifiers. The analysis is based on a set of dynamic signature parameters, such as pen coordinates, pressure, velocity, acceleration, and their derived features. Statistical metrics – mean values, standard deviations, coefficients of variation, entropy, and equivocation – are combined with Discrete Fourier Transform (DFT), Discrete Cosine Transform (DCT), and Discrete Wavelet Transform (DWT) to form an informative feature space. These features are then used to synthesize an MLP classifier whose architecture is adapted to the input data. Results . Experimental results confirm that using secondary features significantly increases identification accuracy compared to traditional methods. A set of 3–5 key parameters along with their spectral derivatives allows for accuracy levels of 0,8 to 0,95 with a limited number of users, maintaining around 0,7 when scaling. The average improvement in identification accuracy was 25–35% over statistical methods and 5–15% over metric-based algorithms. Conclusion . To ensure the required level of identification reliability, it is recommended to apply a multi-level approach involving separate processing of dynamic signature parameters followed by result integration. The most effective configurations were based on neural network models combined with metric and correlation methods operating in the space of spectral and statistical features.