ELIDS: Ensemble Feature Selection for Lightweight IDS against DDoS Attacks in Resource-Constrained IoT Environment

Abstract

Distributed Denial of Service (DDoS) attacks are increasingly prevalent in today’s digital era, particularly due to the growth of unprotected Internet of Things (IoT) devices. This escalation in attacks necessitates the development of adaptive Intrusion Detection Systems (IDS) that are both efficient in detection and less resource intensive, suitable for resource-constraint IoT devices. Feature Selection (FS) techniques are most popular for developing efficient and lightweight IDS, but their reliance on a single method can be risky due to inherent dataset biasness, necessitating diverse and adaptable FS approaches. This paper proposes a new Ensemble FS approach, termed as Ensemble Feature Selection for Lightweight IDS (ELIDS), by leveraging the strength of seven different filter-based methods. ELIDS not only reduces features but also converges the selection towards the most significant features out of those identified by the individual FS methods. Considering several learning algorithms, robust classification models are built out of ELIDS that are comprehensively evaluated for performance and resumption through both in-domain and cross-domain testing. During in-domain testing, results indicate that classification models built out of ELIDS generally exhibit similar accuracy to those built out of the existing individual FS techniques, where peak accuracy rates of 99.8% are achieved. However, cross-domain testing reveals that classifiers built using individual FS methods fall noticeably in their accuracy whereas the proposed ELIDS show robustness by outperforming existing techniques by large margins. In-fact, ELIDS exhibits a higher accuracy rate by at-least 23.7% over existing solutions while considering Random Forest (RF) as the learning algorithm.