INFORMATION SECURITY FOR SOCIAL NETWORK CYBERSPACE BY DETECTING WAYS OF DISINFORMATION SPREAD AND INAUTHENTIVE CHAT USERS' BEHAVIOR BASED ON DEEP LEARNING AND MULTIMODAL ANALYSIS

Abstract

The article considers the problem of identifying sources and methods of disseminating disinformation and inauthentic behaviour among chat users in the modern information space. The relevance of the study stems from the rapid growth of digital content, the spread of social networks and chat platforms, and the intensification of information attacks, manipulative influence, and the use of automated bot networks. It is shown that traditional approaches based exclusively on the analysis of text content have significant limitations due to their inability to fully account for the behavioural characteristics of users, multimedia components, and structural relationships among participants in the information environment. The first section of the work analyses the current state of the problem of disinformation detection and investigates the main methods for automatic analysis of false content and inauthentic user behaviour. Traditional machine learning approaches, modern deep learning models, multimodal analysis methods and graph neural networks are considered. A comparative analysis of the advantages and disadvantages of existing approaches is conducted, problematic aspects are identified, and promising areas for development are outlined. In the problem statement section, a conceptual research pipeline is proposed that covers multimodal data collection and preprocessing, feature extraction, multimodal fusion, training deep learning models, building user interaction graphs, analysing information cascades, and identifying potential sources of disinformation. A mathematical model integrating text, behavioural, and graph characteristics is developed. The methods and materials section describes the mathematical apparatus of the model, which combines Transformer models for text analysis, convolutional neural networks for visual feature extraction, recurrent networks for behavioural analysis, and Graph Neural Networks for modelling the structure of user interactions. A multimodal feature fusion mechanism using attention mechanisms and combined loss functions is proposed. In the experimental part, a study was conducted using multimodal datasets comprising text messages, activity time series, and user network characteristics. The results demonstrated the effectiveness of the proposed approach in identifying potential sources of disinformation, detecting anomalous behavioural patterns, and predicting how information messages spread. Analysis of interactive distribution graphs and probability maps confirmed the system's ability to adapt to different network structures and information cascade types. The practical significance of the study lies in the possibility of using the developed approach in information space monitoring systems, cybersecurity, social networks, chat platforms and early detection systems of information threats.