Towards a robust, effective and resource-efficient machine learning technique for IoT security monitoring.

Zakariyya, Idris

Abstract

Internet of Things (IoT) devices are becoming increasingly popular and an integral part of our everyday lives, making them a lucrative target for attackers. These devices require suitable security mechanisms that enable robust and effective detection of attacks. Machine learning (ML) and its subdivision Deep Learning (DL) methods offer a promise, but they can be computationally expensive in providing better detection for resource-constrained IoT devices. Therefore, this research proposes an optimization method to train ML and DL methods for effective and efficient security monitoring of IoT devices. It first investigates the feasibility of the Light Gradient Boosting Machine (LGBM) for attack detection in IoT environments, proposing an optimization procedure to obtain its effective counterparts. The trained LGBM can successfully discern attacks and regular traffic in various IoT benchmark datasets used in this research. As LGBM is a traditional ML technique, it may be difficult to learn complex network traffic patterns present in IoT datasets. Therefore, we further examine Deep Neural Networks (DNNs), proposing an effective and efficient DNN-based security solution for IoT security monitoring to leverage more resource savings and accurate attack detection. Investigation results are promising, as the proposed optimization method exploits the mini-batch gradient descent with simulated micro-batching in building effective and efficient DNN-based IoT security solutions. Following the success of DNN for effective and efficient attack detection, we further exploit it in the context of adversarial attack resistance. The resulting DNN is more resistant to adversarial samples than its benchmark counterparts and other conventional ML methods. To evaluate the effectiveness of our proposal, we considered on-device learning in federated learning settings, using decentralized edge devices to augment data privacy in resource-constrained environments. To this end, the performance of the method was evaluated against various realistic IoT datasets (e.g. NBaIoT, MNIST) on virtual and realistic testbed set-ups with GB-BXBT-2807 edge-computing-like devices. The experimental results show that the proposed method can reduce memory and time usage by 81% and 22% in the simulated environment of virtual workers compared to its benchmark counterpart. In the realistic testbed scenario, it saves 6% of memory footprints with a reduction of execution time by 15%, while maintaining a better and state-of-the-art accuracy.

Citation

ZAKARIYYA, I. 2022. Towards a robust, effective and resource-efficient machine learning technique for IoT security monitoring. Robert Gordon University, PhD thesis. Hosted on OpenAIR [online]. Available from: https://doi.org/10.48526/rgu-wt-1987917

• Thesis Type: Thesis
• Deposit Date: Jun 15, 2023
• Publicly Available Date: Jun 15, 2023
• DOI: https://doi.org/10.48526/rgu-wt-1987917
• Keywords: Internet of Things (IoT); Network security; Systems security; Cybersecurity; Machine learning
• Public URL: https://rgu-repository.worktribe.com/output/1987917
• Related Public URLs: https://rgu-repository.worktribe.com/output/2015652 (Related journal article)
• Award Date: Dec 31, 2022