A Hybrid Semi-Supervised Learning Approach Based Dimensionality Reduction for Emerging IoT Cyber Attack Analysis

Abstract

With internet connectivity's dynamic development and network communication escalating, an upward trend in advanced threats and cyber attack crimes is being witnessed, targeting sensitive data and critical systems. As the number of connected devices continues to increase, more opportunities are provided for malicious actors to exploit these vulnerabilities. Therefore, the attack surface expanded, particularly within the Internet of Things (IoT), reinforcing the necessity to design a suitable and adaptable security model for protecting data and minimizing the damage caused by network system intrusions and attacks. Machine learning is a swift and flexible way to develop a cybersecurity model. Clustering and other unsupervised learning techniques are frequently used to find hidden patterns or anomalies without needing labeled data. Still, they are limited in accurately distinguishing between normal and malicious behaviors, particularly when dealing with emerging attacks (such as zero-day attacks). This paper provides a hybrid semi-supervised classification-based clustering approach model to distinguish malicious emerging cyber-attacks. Initially, unsupervised clustering methods were implemented to group related data points and find underlying patterns to the reduced data dimensionality under Principal Component Analysis (PCA). Then, the pseudo-labels generated by clustering were used in supervised tasks with a Support Vector Machine (SVM) for classification. The experiments have been carried out on two IoT datasets, namely MQTTset and IoTID20, and the results confirm that the hybrid strategy is a viable and effective approach to dealing with IoT security issues in complicated and real-world situations.