Babu, Rasel Hossain and Alam, Shafi and Hassan, Md Zahid and Nopu, Md Asaduzzaman and Soumik, Md Shadman (2025) DEVELOPMENT OF ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING BASED CYBERSECURITY TECHNOLOGIES FOR DETECTING THREATS, PROTECTING SENSITIVE GOVERNMENT DATA, AND STRENGTHENING THE RESILIENCE OF DIGITAL INFORMATION INFRASTRUCTURE. International Journal of Applied Mathematics, 38 (4s). ISSN 1314-8060

Text Development of Artificial Intelligence and Machine Learning Based Cybersecurity Technologies IJAM copy.pdf Download (723kB)

Abstract

Government agencies that hold sensitive data, and the digital information infrastructure that supports them, face a steady increase in the volume, speed, and variety of cyber threats, and manual analysis can no longer keep pace with the telemetry that modern networks produce. Machine learning (ML) is widely proposed as a way to detect intrusions, malware, and anomalous behavior at scale, yet the literature often reports accuracy figures that do not translate into operational value because they ignore class imbalance, unequal error costs, and the fragility of models under changing conditions. This educational article presents a framework for developing and evaluating AI and ML-based threat detection technologies in settings where sensitive data must be protected and service continuity matters. The framework combines standard classification metrics with a cost-sensitive decision model, a threshold-selection rule derived from Bayes decision theory, and a simple robustness protocol based on feature noise. To demonstrate the framework without using any real organizational data, we generated a synthetic imbalanced dataset of 20,000 records with approximately ten percent malicious samples and compared a logistic regression baseline, a random forest, a gradient boosting classifier, a multilayer perceptron, and an unsupervised isolation forest. In this synthetic setting the multilayer perceptron achieved the highest F1 score (0.792), the random forest achieved the highest precision (0.961), and tuning the decision threshold of the random forest to minimize expected cost reduced that cost by roughly 63 percent relative to the default threshold of 0.5. Performance of all supervised models degraded as input noise increased. These results should not be generalized to real traffic; their purpose is to show how the evaluation pipeline works and what questions practitioners should ask of any reported detector.

Item Type:	Article
Uncontrolled Keywords:	artificial intelligence, machine learning, cybersecurity, threat detection, government data protection, intrusion detection, class imbalance, cost-sensitive learning, anomaly detection, resilience
Subjects:	H Social Sciences
Depositing User:	admin eprints
Date Deposited:	19 Oct 2026 01:23
Last Modified:	19 Oct 2026 01:23
URI:	http://eprints.umsida.ac.id/id/eprint/17218

Actions (login required)

View Item