Dynamic Signature Verification Method Based on XGBoost Using Micromotor Features and Adaptive Decision-Making Threshold

Abstract

The relevance of the study is driven by the need to create highly reliable biometric systems resistant to unauthorized access and high-quality signature forgeries in the context of global business process digitalization. Traditional static verification methods are gradually losing effectiveness due to the emergence of new tools for reproducing graphical patterns. In this regard, the author conducted a comprehensive comparative analysis of a custom-developed ML solution against the recognized commercial industry standard — Wacom Ink SDK for Verification. The primary focus is placed on the systems' ability to differentiate a genuine signature from a skilled forgery performed by an attacker with visual access to the original.

The experimental study was conducted using a professional Wacom STU-540 signature pad, which provides a high sampling rate (200 Hz) and 1024 levels of pressure sensitivity. The choice of such hardware enabled the collection of detailed time series, including pen coordinates (x, y), tilt angles, azimuth, and the dynamics of pressure changes. Based on the analysis of real-world data from six users, who provided both reference samples and imitation forgery attempts, a multi-criteria optimization of the decision-making threshold was performed. It was established that at an adaptive threshold value, the developed system demonstrates balanced performance metrics: a False Acceptance Rate (FAR) of 1.33 % and a False Rejection Rate (FRR) of 6.67 %.

The scientific novelty of the work lies in the substantiation and selection of a set of the most informative dynamic parameters for the XGBoost gradient boosting model. Unlike standard approaches, the proposed model considers not only the signature geometry but also derivative characteristics: instantaneous velocity, acceleration, and pressure change rate (jittering). The use of the regularization mechanism in XGBoost prevented overfitting on small datasets, which is typical for single-user biometric data. This allowed for achieving an accuracy level competitive with closed-source commercial algorithms (Equal Error Rate EER ~4 %).

Features of the software implementation include a data preprocessing stage where interpolation is applied to align time steps and feature normalization is used to bring them to a single scale. This is critical for the stable operation of the XGBoost model, as each user's hand micromotors possess unique amplitude characteristics. The developed processing pipeline enables real-time verification, spending less than 100 ms per request, making the system suitable for high-intensity document management environments.

The practical significance of the results lies in the possibility of integrating the proposed approach into corporate electronic document management systems and automated personnel identification systems (e.g., for verifying drivers or financial managers). The main advantage is the lack of dependence on expensive proprietary software and licensing restrictions while maintaining a high level of security.

The comparison results confirmed that the use of gradient boosting, taking into account deep micromotor parameters, effectively detects intruders even during attempts of meticulous visual style copying.