RESEARCH INTO MECHANISMS FOR DETECTING SOURCES AND WAYS OF FAKE NEWS AND PROPAGANDA DISSEMINATION IN SOCIAL NETWORKS CYBERSPACE

Abstract

The article presents an information technology for detecting sources and routes of fake news and propaganda distribution in social networks and online media. The proposed approach is based on vectorization of text publications using language models, calculation of cosine similarity and clustering of messages to identify thematically similar groups. Based on the formed clusters, a chronological graph visualization is built, which allows identifying the primary sources of disinformation, key relays and the speed of its distribution. The experimental study was conducted on a combined dataset of over 2,000 posts, in which the shares of true and fake messages are relatively balanced. Interaction analysis showed that about 75% of all posts receive no more than 20 likes, while only less than 5% of posts form a “long tail” with hundreds and thousands of reactions. At the same time, disinformation posts are more likely to either remain almost unnoticed or to sharply gain abnormally high popularity in a short period of time. Analysis of distributions revealed that approximately 80% of posts have no more than 5 reposts, however, the share of fakes among the most widely shared messages is growing significantly. Platform analysis showed that reliable content prevails on web resources and authoritative media, while on social networks and messengers (in particular Telegram) the ratio of true and fake messages is close to equilibrium, and in some network’s disinformation dominates. A comparison of embedding models showed that the OpenAI model provides a clearer separation of messages in the feature space and allows detecting up to 10–11 significant clusters at an optimal cosine similarity threshold of τ ≈ 0.55–0.60, while for the local model the optimal threshold is much higher (τ ≈ 0.86–0.88). The constructed propagation graphs showed that the time interval between the initial publication of a fake and its appearance on other platforms can be from several hours to several days, and the transformation of reposts into “original” publications is a typical mechanism for hiding the source.