Electronic Thesis and Dissertation Repository

Thesis Format

Monograph

Degree

Master of Science

Program

Computer Science

Collaborative Specialization

Artificial Intelligence

Supervisor

Fadlullah, Zubair

Abstract

As Internet of things (IoT) networks expand and more real-time data are managed by Software Defined networking (SDN) controllers, the risk of cyber threats grows, making it crucial to ensure security and efficiency. Traditional Machine Learning approaches with centralized data processing struggle with the scalability needs of modern IoT systems. Federated Learning (FL) has emerged as a solution, enabling model training across devices without centralizing data, addressing privacy concerns and reducing server load. However, IoT networks face challenges like inconsistent data across devices, which can hinder FL's effectiveness in detecting IoT attacks. Additionally, the computing demands of model training can strain SDN controllers, increasing latency. This study addresses these challenges by (1) evaluating FL methods, including FedAvg, FedProx, and Scaffold, under varied data distributions to address statistical heterogeneity in detecting IoT attacks, and (2) proposing Resampling Dynamic Update Strategy or REDUS, a resampling technique that prioritizes challenging data while reducing redundancy, thereby lowering training time and latency. Experimental results show that REDUS achieves a 72.6% training time reduction with only 1.62% accuracy loss, providing an efficient solution for IoT attack detection in resource-constrained settings.

Summary for Lay Audience

The IoT paradigm offers connectivity to everyday devices like home security systems, wearable fitness trackers, and smart appliances, allowing them to collect and share data. This connectivity has dramatically transformed how we live; however, these IoT devices are frequently exposed to cyber threats, since more devices mean more opportunities for attackers. These threats include cyber attacks that can disrupt devices, steal private information, or compromise entire networks. Securing IoT networks is challenging because they generate massive amounts of data and often lack the power and resources to handle complex security measures. The research in this thesis focuses on protecting IoT networks via Federated Learning dubbed FL. Unlike traditional methods that gather data in a central location, FL allows each device to train on its own data and then share only the training results, not the data itself. This approach retains sensitive information on individual devices, reducing privacy risks. FL, however, is not a perfect solution; since data across devices vary widely, it can make the model training less accurate. Additionally, IoT networks often share resources with Software-Defined Networking (SDN), which manages network traffic and keeps systems function smoothly.

To address these issues, this research explores two main approaches. The first approach improves FL’s accuracy in detecting cyber threats across different devices with varied data, making security systems more reliable. The second approach introduces an “adaptive resampling” technique, which trains AI models faster by focusing only on challenging data points, reducing the processing time and energy needed for training without compromising accuracy. Together, these methods establish a scalable, privacy-focused, and efficient framework to enhance IoT security without significantly impacting network performance, providing a practical solution for deploying AI in IoT environments, which demand both security and resource efficiency.

Creative Commons License

Creative Commons Attribution 4.0 License
This work is licensed under a Creative Commons Attribution 4.0 License.

Available for download on Thursday, July 30, 2026

Share

COinS