Hybrid Intelligence Model for Malicious URL Detection Using Machine Learning and Graph Neural Networks

Abstract

The rapid expansion of interconnected cyber-physical systems and IoT-enabled Industrial Internet of Communications (IICs) has substantially heightened the risk of sophisticated cyber threats. Intrusion Detection Systems (IDSs) serve as a crucial defensive mechanism for safeguarding network security; however, conventional IDS methods often exhibit limited detection accuracy, elevated false-positive rates, and insufficient adaptability to novel and evolving attack patterns. To address these challenges, this paper presents a novel deep learning–based IDS framework designed to efficiently detect vulnerabilities and malicious activities in cyber-physical environments. The proposed framework combines unsupervised learning techniques with deep discriminative models and leverages a Generative Adversarial Network (GAN) to improve detection performance under complex network traffic conditions. Comprehensive evaluations were carried out using benchmark datasets, including NSL-KDD, KDDCup99, and UNSW-NB15, demonstrating the model’s effectiveness. The results indicate substantial performance gains, achieving high accuracy and enhanced reliability with a dropout rate of 0.2 over 25 training epochs. Additionally, the proposed model attained the highest True Negative Rate (TNR) and Detection Rate for multiple attack categories, such as BruteForce-XSS, BruteForce-WEB, DoS Hulk, and DoS LOIC HTTP