DEVELOPING MACHINE LEARNING ALGORITHMS TO COMBAT NETWORK INTRUSIONS
International Journal on Technical and Physical Problems of Engineering, vol.17, no.64, pp.450-459, 2025 (Scopus)
- Publication Type: Article / Article
- Volume: 17 Issue: 64
- Publication Date: 2025
- Journal Name: International Journal on Technical and Physical Problems of Engineering
- Journal Indexes: Scopus
- Page Numbers: pp.450-459
- Keywords: Anomaly Detection, Cybersecurity, Intrusion Detection, Machine Learning, Network Attacks
- Open Archive Collection: Article
- Azerbaijan State University of Economics (UNEC) Affiliated: No
Abstract
Modern information systems are under continuous threat from increasingly sophisticated network attacks, making the deployment of adaptive and intelligent defense mechanisms a critical priority for ensuring cybersecurity. To address this issue, the present work proposes a combined ensemble strategy for network intrusion detection, merging traditional supervised learning models with unsupervised anomaly recognition methods in order to improve the accuracy of detecting and preventing harmful activities. The proposed architecture combines Gradient Boosting Classifiers, Random Forest models, and Deep Neural Networks within a unified decision fusion framework. This design aims to leverage the complementary strengths of each algorithm-robust feature handling, high classification accuracy, and nonlinear pattern recognition-to improve overall detection performance. The efficiency of the suggested approach was assessed through experiments carried out on two well-established benchmark datasets commonly employed in intrusion detection studies: CIC-IDS2017 and UNSW-NB15. The datasets were preprocessed and subjected to feature selection using mutual information criteria before training the models. A benchmark comparison indicated that the proposed ensemble approach provided improved detection rates, precision, and recall over conventional classifiers, with accuracy improvements ranging from 4% to 7% over the best-performing individual models. The findings further indicated a notable decrease in false alarm rates, thereby improving the overall dependability of the detection framework in practical network settings. This research underscores the promise of integrating different machine learning approaches for building more resilient and adaptive intrusion detection systems. The proposed approach can be extended for real-time applications and adapted to emerging attack patterns, contributing to the development of next-generation cybersecurity solutions.