Ensemble Feature Selection based Lightweight IDS Tailored for DDoS Attacks Detection over IoT Devices

• Published in: 2024 International Visualization, Informatics and Technology Conference (IVIT) pp. 20 - 27
• Main Authors: Fatima, Mahawish, Rehman, Osama, Rahman, Ibrahim M. H
• Format: Conference Proceeding
• Language: English
• Published: IEEE 07.08.2024
• Subjects: Feature Selection; Internet of Things; Intrusion Detection; Lightweight; Machine Learning
• DOI: 10.1109/IVIT62102.2024.10692650

The attack surface has grown due to the widespread use of weak Internet of Things (IoT) devices, and distributed denial of service (DDoS) attacks are becoming more common. The necessity for adaptive intrusion detection systems (IDS) that are effective at detecting threats and need less resources is highlighted by the increase in attacks, particularly for IoT devices with limited resources. Although feature selection (FS) techniques are widely used for creating lightweight intrusion detection systems (IDS), depending solely on one approach can be dangerous and lead to bias in the dataset. Therefore, a flexible and varied FS approach is required. ELIDS (Ensemble Feature Selection for Lightweight IDS), an Ensemble FS strategy that utilizes the advantages of seven different filter-based techniques, is presented in this work. The primary goal of ELIDS is to choose the most important features found by each FS approach. This leads to build strong classification models, which are then thoroughly assessed for both performance and resource efficiency using both in-domain and cross-domain testing. The evaluation findings demonstrate that the suggested model reaches a peak accuracy of 100% for in-domain testing in addition to being lightweight. Cross-domain testing, however, shows that classifiers constructed using individual FS methods suffer a notable reduction in accuracy, while ELIDS-based classifiers exhibit robustness and significantly exceed current methods. Specifically, ELIDS-based classifiers outperform other models by 24%, particularly when Random Forest (RF) is used as the learning technique.