PHISHING DETECTION IN ELECTRONIC COMMUNICATION CHANNELS

Abstract

The article is devoted to increasing the effectiveness of information protection tools in domestic cyberspace by automated detection of phishing in text messages of electronic communication channels. It is shown that phishing remains one of the dominant vectors of cyberattacks, in particular in conditions of active use of social engineering methods and psycho-emotional influence, which significantly complicates its timely detection by traditional signature and statistical methods. It is established that most of the known neural network solutions are characterized by high resource intensity, require the formation of significant volumes of marked-up training data and are insufficiently adapted to the specifics of domestic content, which is characterized by limited marked-up corpora, morphological variability and sensitivity to contextual social engineering influences. To overcome these limitations, the article proposes a method for detecting phishing, which is based on the use of proven dialogical intelligent assistants based on large language models in the mode of formalized dialogical interaction. The method involves automated analysis of text messages by submitting standardized queries, formed taking into account typical features of phishing attacks. A classification of phishing messages by a vector of detection features is proposed. For each type, target query predicates are formed, which provide deterministic and interpretable analysis of the results. The developed method involves preprocessing of the text, formation and submission of queries to dialogic intelligent assistants, aggregation of responses, and assessment of the presence of phishing using the weighted linear convolution mechanism. Experimental studies have shown that the average classification accuracy when using the proposed method is 97.5%, which corresponds to the level of the best known solutions of a similar purpose. At the same time, the implementation of the proposed method does not require resource-intensive training of neural networks, the formation of large volumes of labeled data, and the creation of specialized hardware and software, which ensures the efficiency of creating an effective phishing detection system in domestic electronic communication channels.