Method of Biometric Authentication of Personnel of Critical Infrastructure Facilities by Facial Image and Iris of the Eye Using Neural Network Tools

The widespread implementation of neural network biometric authentication tools based on facial and iris images at critical infrastructure facilities has made it possible to increase the level of security and efficiency of personnel identification. At the same time, modern requirements dictate the need to increase resistance to spoofing attacks, adaptability to interference from the real environment, as well as expand functionality to take into account the psycho-emotional state of personnel at the time of authentication. Traditional neural network methods based on neural networks with monolithic architecture are limited in implementing these requirements due to insufficient flexibility and difficulty in adapting to variable video recording conditions. Therefore, this article proposes a method for biometric authentication of personnel at critical infrastructure facilities based on facial and iris images using neural network tools. The method implements multi-step adaptive processing of the video stream and comprehensive execution of procedures necessary for effective authentication in conditions of variability of image parameters. The stages of the method are structured as a sequence of procedures that correspond to typical biometric authentication tasks, each of which is implemented by a separate neural network. For the first time, a multi-level analysis of signs of spoofing attacks is provided within a single architecture, taking into account visual facial artifacts, artifacts of the external environment, the dynamics of survivability parameters and emotional state. A mechanism for automatic evaluation of display indicators, dynamic formation of a set of images suitable for processing and their pre-processing to improve quality is proposed. Interference filtering is carried out by semantic segmentation and element-wise masking of pixel areas, which allows excluding areas with shadows, overlaps or extraneous objects from processing. A key feature is the targeted use of available and tested pre-trained neural network models with open source (FaceMeshV2, Visual Transformer, DeepPixBiS, Siamese Network, Two_channel Networks, MobileNet), adapted to the conditions of video recording at critical infrastructure facilities. According to preliminary estimates, the use of adaptive video stream processing and multi-level analysis of spoofing attacks allows to increase the accuracy of spoofing detection by 25–35% and reduce the development time of the system based on open models by at least 1.5 times.