Detecting Zero-Day Attacks under Concept Drift: An Online Unsupervised Threat Detection System

In recent years, there has been significant interest towards mechanisms for detecting cyber-security threats. However, the dynamic nature of modern systems and networks poses significant challenges for threat detection systems exploiting machine learning models, since shifts in data' s statistical distribution over time, known as concept drift, can cause severe performance degradation. In this scenario, traditional static systems often need manual retraining by human operators, leaving networks exposed to vulnerabilities in the interim. Moreover, the challenge of detecting zero-day attacks through semi-supervised or unsupervised models remains a critical aspect that has garnered much attention in the literature. This work introduces an unsupervised online threat detection system designed to identify anomalous traffic indicative of zero-day attacks, while explicitly handling concept drift by automating retraining processes only when necessary. An extensive experimental evaluation on the real-world IoT-23 dataset, encompassing network traffic from IoT devices and malicious traffic from malware-infected devices, showcases the system' s efficacy, showing superior performance in real-time threat detection compared to traditional static approaches.