A method for detecting financial phishing in email based on a judicial interaction model of intelligent agents

Abstract

Relevance: stems from the need to improve the detection of financial phishing in email, in particular when fraudulent messages use sophisticated social engineering techniques and contain no obvious technical indicators of compromise. Traditional filtering tools and solutions based on individual large language models may demonstrate insufficient effectiveness in detecting such phishing because they typically evaluate the message text in isolation and do not compare different interpretations of its content. The aim of the article is to develop a method for detecting financial phishing in email based on a model of judicial interaction among intelligent agents, which is intended to improve the effectiveness of its recognition. Objectives: to develop a model of a multi-agent environment of judicial interaction among intelligent agents with assessments represented as fuzzy numbers; to develop a method for detecting financial phishing in email that implements the proposed model as sequential stages of analyzing the incoming message and aggregating the fuzzy assessments of the agents; and to evaluate the effectiveness of the method experimentally. Methods: the study applies a multi-agent model of judicial interaction among intelligent agents based on large language models, together with fuzzy logic. Individual verdicts are expressed as linguistic assessments mapped to triangular fuzzy numbers and aggregated using competence weighting coefficients. Scientific novelty lies in adapting the adversarial procedure of judicial proceedings to the organization of interaction among intelligent agents, which involves the formation of opposing reasoned positions of the prosecution and the defense, their evaluation by a panel of judges, and the aggregation of the judges’ verdicts based on fuzzy logic, thereby making it possible to improve the accuracy of financial phishing recognition. Practical significance: the proposed method can be used in email protection tools where accuracy takes priority over resource consumption, providing a textual justification of the message classification decision. Results: experimental studies on a corpus formed by the authors, using Claude, ChatGPT, Gemini, and DeepSeek, showed a lower classification error rate for the proposed method than for individual large language models and for the weighted integration method without judicial interaction, under the same experimental protocol. Conclusions: the proposed method makes it possible to improve the accuracy of financial phishing recognition in email on the basis of judicial interaction among intelligent agents implemented using large language models, without additional training of these models.