Lightweight Meta-Learning BotNet Attack Detection

Modern society is increasingly dependent on numerous Internet of Things (IoT) devices to assist in a variety of scenarios, such as smart homes and cities, healthcare systems, and cyber-physical systems. Despite IoT's increasing popularity, IoT security remains a challenge due to the multitude of attack vectors. Existing cyber-attack defense methods attempt to protect the network from both within and outside the network. Network intrusion detection systems (NIDSs) act as device borders within network security and offer a potential defense methodology. This research analyzes the performance of an Artificial Intelligent Internet of Things (AIoT) lightweight botnet attack detection model by deploying meta-learning ensemble botnet detection models and evaluates the capability of a single-board system in addressing cyber-attack threats. The Aposemat IoT-23 [1], UC Irvine KDD99 [2], and UNSW TON [3] data sets provide IoT and network traffic network flow captures which are used to evaluate the proposed meta-learning methodologies. Experiments show that deployment of our proposed methodologies on edge devices exhibits similar results to PC-based Desktop CPU-trained models. Over the three data sets, when considering a binary classifier (benign versus malignant), our models can consistently achieve above 97.9% accuracy with a false positive rate (FPR) less than 3.8% and an inference time less than 3.95 s. In this work, we show that for binary classification our meta-learners provide consistently stable high accuracy low FPR performance across all three data sets, while maintaining reasonable inference times.