Розробка методу антиспуфінгу зображень в системах біометричної безпеки з використанням ML

• Date: 2025
• Main Authors: Палагін, Володимир, Івченко, Олександр, Палагіна, Олена, Філіпов, Віталій, Гавриш, Олександр, Байрак, Анатолій, Пташкін, Роман
• Format: Article
• Language: Ukrainian
• Published: Kamianets-Podilskyi National Ivan Ohiienko University 2025
• Online Access: http://mcm-tech.kpnu.edu.ua/article/view/332637
• Journal Title: Mathematical and computer modelling. Series: Technical sciences

This article addresses the problem of detecting presentation attacks (spoofing) in facial biometric authentication systems. Given the growing prevalence of spoofing attacks using printed photos, video replays, and 3D masks, there is an urgent need to develop robust anti-spoofing mechanisms. The objective of this study is to develop a multi-level combined method for spoofing detection that integrates both physical and behavioral facial features using machine learning models for adaptive decision-making. The proposed system includes four functional modules: edge detection, motion analysis, blink detection, and smile detection. Each module generates a binary decision regarding the presence of liveness indicators, which are subsequently passed to an integration block. Unlike traditional approaches with fixed weighting schemes, this system computes weights adaptively based on a trained machine learning model. This enables the dynamic adjustment of each module's influence depending on environmental conditions, video quality, and individual facial characteristics. The scientific novelty of this work lies in the development of a flexible mechanism for optimizing weighted coefficients of anti-spoofing modules based on the outcome of the learning phase. The system demonstrates the ability to self-adjust, enhancing overall detection accuracy while reducing false acceptance/rejection rates in challenging scenarios. The proposed model dynamically balances the contributions of physical and behavioral features to the final decision in real time. The system is implemented as a modular architecture and tested on a controlled dataset containing various spoofing scenarios. Experimental results demonstrate high detection accuracy (up to 100% in test settings) and resilience to variability in input data. The article presents the operational algorithm, mathematical formulation of the modules, the principle of decision integration, and scalability potential. The proposed solution shows promise for practical implementation in mobile devices, access control systems, and online identity verification platforms.