Abstract
This paper presents the research and development of an intelligent hybrid model based on deep learning methods for malware detection in information and communication infrastructures. The goal of the study is to create an architecture combining convolutional (CNN), recurrent (GRU), and graph (GNN) neural networks with ontology-based feature normalization. This approach provides a comprehensive analysis of the static, behavioral, and structural characteristics of malware, increasing resilience to zero-day attacks and polymorphic threats. The EMBER-2018, Malimg, CICAndMal2017, and CICMalDroid-2020 open datasets were used for experimental validation. The experiments showed that the proposed hybrid architecture provides a classification accuracy of 98.7% and a ROC-AUC of 0.99, outperforming isolated deep learning models and traditional machine analysis methods. The introduction of an ontological knowledge model increased the system's interpretability and reduced the false-positive rate to 1.7%. The developed software prototype demonstrated the solution's applicability in real-time systems (SOC/IDS) and demonstrated potential for further scaling and implementation in practical cybersecurity systems. The scientific novelty of the study lies in the integration of multimodal analysis and ontological modeling, which enables more accurate and explainable detection of malicious objects in complex digital environments.