DETECTION OF MANIPULATIVE COMPONENT IN TEXT MESSAGES OF MASS MEDIA IN THE CONTEXT OF PROTECTION OF DOMESTIC CYBERSPACE

Abstract

The problem of the article is to increase the effectiveness of information security means within the national cyberspace by automated detection of the manipulative component in text messages of the mass media. It is shown that one of the main directions of increasing the effectiveness of such means is the use of large language models, which are capable of performing a deep contextual analysis of a natural language text, taking into account emotional, rhetorical and semantic content. It is established that most of the known solutions in the field of detecting text manipulations using large language models are quite difficult to adapt to the conditions of practical application in systems for protecting domestic cyberspace due to the need to create a powerful hardware and software infrastructure to service the corresponding LLM-means and the need to form specialized training samples that take into account the main types of manipulative influences characteristic of the realities of information confrontation. To overcome these limitations, the article proposes a concept for using LLM tools (GPT, Gemini, DeepSeek, Grok, etc.), which is based on the implementation of dialogic interaction with pre-standardized formalized queries that take into account the main types of manipulative influences. Such influences, according to the classification, include emotional-manipulative messages, information substitutions, discrediting narratives, context manipulation, propaganda constructs, exploitation of socially sensitive topics, and artificial formation of public opinion through bot activity. A method of interaction with LLM has been developed, which includes the stages of text pre-processing, the formation of queries at the basic, typological, and interpretative levels, as well as the interpretation of LLM responses in a formalized form. Experimental studies involving three groups of texts (scientific, political, and destructive-propaganda) have shown that the analysis results obtained using the GPT-4-turbo LLM tool agree with expert assessments by an average of 87%, which indicates a high level of reliability of the results obtained and confirms the practical effectiveness of the proposed solutions. It is shown that it is possible to increase the accuracy and stability of assessments of the manipulative component by retraining the LLM tools by including examples of correct answers in the query, which allows to increase the consistency of the results without the need for complete retraining of the model.