IoT Based Resource Constrained Intrusion Detection and Classification

Abstract

With the rapid increase in today’s IoT devices variety, availability, and usage securing the network is crucial. Intrusion Detection Systems (IDS) are considered the next line of defense when it comes to security. The main three types of IDS in IoT are signature-base, anomaly-base, and hybrid IDS. Many limitations are still persistent when detecting intrusion in IoT network including the resource-constrains, high false positives rates, and privacy preservation. Researches advances in the detection of various and new attacks are necessary. There are few publicly available data sets that simulates reality that are used to allow advancement in the detection of intrusion and attacks. In this research we investigate using the BoTNeTIoT-L01 dataset the efficiency of various machine learning models in both centralized and federated learning framework for the detection and classification of botnet attacks. All models were tested on resource constraint Raspberry Pis connected via Transmission Control Protocol (TCP). Performance metrics (eg. accuracy, precision, recall, ..etc.) are reported as well as resource consumption (eg. CPU usage, RAM usage, CPU temperature, ...etc.). The results showcased that Decision Tree achieved over 99% accuracy in both detection and classification within the centralized and federated learning framework outperforming recent research. In addition, it maintained a comparatively low resource consumption of with around 25%. Many enhancement could be applied to this research including the exploration of the effects of automated hyperparameter tuning, and Generative Adversarial Neural Network to create synthesized data and offer a balanced dataset. Additionally, this research could be extended by exploring real-time detection and classification on resource-constrained environments, applying transfer learning to adapt models to new datasets and attack scenarios, strengthening model robustness against adversarial attacks such as label flipping and poisoning, and incorporating explainable AI methods like SHAP and LIME to improve transparency and interpretability of model decisions.

Date of Award	7 May 2025
Original language	American English
Supervisor	U Zeyar Aung (Supervisor)