Artificial Intelligence-Based Intrusion Detection Systems for Advanced Cyber Threat Detection: A Machine Learning Approach

Abstract

The exponential growth of network connectivity, cloud computing, and the Internet of Things (IoT) has dramatically expanded the attack surface available to malicious actors, rendering conventional signature-based intrusion detection systems (IDS) increasingly inadequate against novel and polymorphic cyber threats. This paper presents an artificial intelligence-based intrusion detection framework that leverages supervised and ensemble machine learning algorithms to detect advanced cyber threats with high accuracy and low false-positive rates. Using the benchmark NSL-KDD and CICIDS-2017 datasets, we systematically evaluate five classifiers—Decision Tree, Support Vector Machine (SVM), Random Forest, Gradient Boosting, and a Deep Neural Network (DNN)—across a unified preprocessing and feature-selection pipeline. Feature dimensionality was reduced using a hybrid information-gain and recursive feature elimination strategy, retaining the most discriminative attributes while lowering computational overhead. Experimental results demonstrate that the Random Forest and Deep Neural Network models achieved detection accuracies of 99.2% and 99.4% respectively, substantially outperforming the baseline Decision Tree classifier. The proposed framework further reduced the false-positive rate to below 1.1% while maintaining sub-millisecond per-flow inference latency, making it viable for near real-time deployment. The findings confirm that AI-driven detection meaningfully improves the identification of Denial-of-Service, probing, and infiltration attacks relative to traditional approaches. This study contributes a reproducible, comparative benchmark of machine learning classifiers for intrusion detection and offers practical guidance on model selection, feature engineering, and class-imbalance mitigation for cybersecurity practitioners.