The dissertation discusses the issues of building user verification systems based on biometric and behavioral data using deep learning neural networks. The need for reliable and highly efficient verification systems is extremely relevant in the fields of security, cybersecurity, personal data protection, medicine, and risk management. Continuous biometric and behavioral signals allow for the implementation of continuous and implicit authentication systems. Since biometric signals are very complex by nature, developing a high-precision verification system requires building new powerful models that have high predictive power and can find deep patterns in data with a complex and deep structure. The goal of the research is to develop and analyze machine learning methods, particularly deep learning neural networks, for user verification based on biometric and behavioral characteristics.
In the dissertation, the following new scientific results were obtained for the first time:
1. New hybrid architectures based on compressive and variational autoencoders using transformers were developed to solve user verification tasks based on behavioral and biometric characteristics, which allowed for improvement in efficiency criteria compared to existing methods.
2. Based on the developed new hybrid architectures, a user verification decision support system was created.
3. A new approach for improving the accuracy of biometric verification systems, based on the use of fractal dimension magnitudes, was developed.
4. Applied scenarios and components of the verification system based on a refined practical methodology for building deep learning systems based on the proposed architectures were identified and further developed.
The theoretical significance of the obtained results lies in the improvement and further development of the methodology for building verification systems based on deep learning neural networks. The created hybrid neural networks allow for a significant increase in the efficiency of biometric verification systems, due to the combination of advantages of components from different architectures in one neural network. Based on the new developed neural network architectures, the impact of fractal dimension magnitudes on the quality metrics of verification systems was discovered and quantitatively assessed.
The practical value of the dissertation work:
1. An original continuous biometric verification user support system based on new hybrid neural network architectures using fractal dimension magnitudes was developed;
2. The developed architectures and refined methodology were implemented in the educational process in the form of the corresponding syllabus, lecture materials, and a training manual-practicum.
The proposed new hybrid architecture, based on compressive autoencoders using transformers, shows a 31% faster inference time and on average 11% lower equal error rate values.
An analysis of the impact of fractal dimension magnitudes on verification systems based on autoencoders was conducted. The positive impact of fractal dimension on the main quality metrics, specifically an average of 13% lower equal error rate and 2.2% higher area under the curve value compared to the method without fractal dimension, was proven.
An automated system for continuous biometric user verification has been proposed, based on newly developed hybrid architectures and taking into account the fractal dimension of the data. The system receives input data from a variety of sensors, which characterize the corresponding biometric or behavioral indicators of a person. During the initialization phase, an initial necessary amount of data is collected for training the new hybrid architectures. Based on an improved practical methodology for setting verification system parameters, appropriate values for the dimensions of input data are selected depending on the characteristics of the sensors before training; hyper-parameters of the deep learning neural network architecture are adjusted; and the fractal dimension of data for each type of sensor is calculated. After training, each model yields values for the relevant criteria (inference time, verification threshold value). Depending on the availability of signals for inference, the system selects the model that covers the broadest context and does not exceed the established permissible value for inference time.
A comparative analysis of different types of autoencoders with classic machine learning methods, was conducted. A significant advantage of autoencoders compared to classic machine learning methods was shown. A deep analysis of the impact of various components of the biometric signal and their quantitative impact on the efficiency of the biometric verification system of the user was conducted and was shown that individual components significantly vary in their impact on quality metrics and demonstrate the effectiveness of combining components to achieve higher accuracy.
