A Novel Dimension Reduction Scheme for Intrusion Detection Systems in IoT Environments

TL;DR

This paper introduces a new distributed autoencoder-based dimension reduction scheme for IoT intrusion detection systems, significantly reducing data size and resource usage while maintaining detection accuracy.

Contribution

It presents a novel autoencoder design that enables efficient, resource-friendly data compression for IoT IDS, addressing bandwidth and resource constraints.

Findings

Achieves about 90% data compression with minimal accuracy loss

Validates effectiveness on three well-known datasets

Reduces bandwidth and computational overhead in IoT IDS

Abstract

Internet of Things (IoT) brings new challenges to the security solutions of computer networks. So far, intrusion detection system (IDS) is one of the effective security tools, but the vast amount of data that is generated by heterogeneous protocols and "things" alongside the constrained resources of the hosts, make some of the present IDS schemes defeated. To grant IDSs the ability of working in the IoT environments, in this paper, we propose a new distributed dimension reduction scheme which addresses the limited resources challenge. A novel autoencoder (AE) designed, and it learns to generate a latent space. Then, the constrained hosts/probes use the generated weights to lower the dimension with a single operation. The compressed data is transferred to a central IDS server to verify the traffic type. This scheme aims to lower the needed bandwidth to transfer data by compressing it and also reduce the overhead of the compression task in the hosts. The proposed scheme is evaluated on three well-known network traffic datasets (UNSW-NB15, TON\_IoT20 and NSL-KDD), and the results show that we can have a 3-dimensional latent space (about 90\% compression) without any remarkable fall in IDS detection accuracy.