Analysis of the effectiveness of neural network architectures for protecting industrial systems from targeted social engineering attacks

This study presents a comprehensive comparative analysis of the effectiveness of modern neural network architectures for countering targeted social engineering attacks on industrial systems. The paper characterizes the main social engineering methods, based on which a group of them is identified that have a critical impact on domestic production. The experimental part of the study is based on an open dataset containing 651 191 URLs categorized into four types: safe resources, defaced links, phishing resources, and malware distributors. The paper presents a systematic evaluation of both classical and innovative machine learning approaches, including Kolmogorov-Arnold networks (KAN), graph neural networks (GNN), capsule neural networks (CapsNets), and their hybrid combinations. The results demonstrate significant superiority of hybrid architectures, where the combination of CNN + LSTM achieved a maximum accuracy of 92.29%, and CNN + KAN demonstrated a result of 92.00%. A detailed analysis revealed the specific effectiveness of various architectures for specific threat categories: CapsNets demonstrated the best results in identifying safe resources (98.60%), while CNN + LSTM were most effective in detecting phishing attacks (72.76%). The scientific novelty of this work lies in establishing a correlation between the type of neural network architecture and the nature of a potential cyberthreat, which creates a methodological basis for developing next-generation adaptive security systems for industrial infrastructure.