METHOD FOR SYNTHESIZING A DISTRIBUTED COMPUTER SYSTEM RESILIENT TO SOCIAL ENGINEERING ATTACKS

Abstract

The paper presents a method for synthesising system architecture for distributed IT – infrastructure systems resistant to social engineering attacks. The core idea is a hierarchical multi-agent design in which a high-level decision-making agent coordinates modality-specialized service agents that collect and verify evidence across textual, structural, and contextual signals.

The architecture integrates several components that aggregates indicators over time, a simulation engine that models user and adversary behavior for training under uncertainty, and an interaction manager that plans information-gathering dialogs with knowledge-graph priors that encode the likelihood of observing particular indicators given an attack type. Learning is guided by reinforcement signals that reward reductions in predictive uncertainty (entropy) and penalize repeated or low-information actions, enabling short, informative interaction sequences and calibrated decisions.

 The approach was instantiated on a combined corpus of emails and web pages labeled with a comprehensive taxonomy of social-engineering tactics and demonstrated that the hierarchical design improves both binary detection and fine-grained tactic/type classification compared with single-agent and other non-hierarchical baselines. In the experiments the proposed system achieves higher accuracy while requiring fewer actions per example, lowers the false-positive rate at stringent operating points, and exhibits improved transfer across modalities through principled coordination of specialized agents. The method provides a principled path to engineering multi-level, adaptable defenses against evolving social-engineering threats in operational environments.