Lightweight IDS for IoT Based on Counter-Propagation Networks

Abstract

The ubiquitous integration of Internet of Things (IoT) networks into almost every aspect of modern life is leading to a multitude of innovative services and applications. However, this massive proliferation of IoT networks exposes them to increasingly complex and massive cyberattacks. Therefore, the development of robust intrusion detection systems (IDS) that can protect these networks from malicious attacks is one of the cornerstones to ensure the sustainable proliferation of these networks. While deep learning-based IDS models have shown remarkable performance, their resource-intensive requirements stand in the way of their deployment on the tiny resources of IoT devices. To bridge this gap, this work proposes a novel IDS solution that synergistically integrates the unsupervised learning paradigm in the form of Growing Hierarchical Self-Organizing Maps (GHSOM) with the supervised learning paradigm in the form of Grossberg-outstar layer to form an improved variant of Counter Propagation Neural Networks. This integration provides the model with a versatile hypothesis space that can be adapted according to the intrinsic patterns and relationships of the IDS dataset in an autonomous and dynamic manner. Moreover, the inherent resilience of the model's components to noise, outliers, concept drift, and catastrophic forgetting enhances its learning robustness and generalization ability. The simplicity and high interpretability of the underlying structures of the datasets are other important advantages of the proposed model. The proposed model has been thoroughly evaluated on several contemporary IoT-IDS datasets from various perspectives. These evaluations demonstrate the ability of the proposed model to achieve near-unity prediction accuracy across all benchmark datasets while consuming only a fraction of the resources required by comparable state-of-the-art models.