Abstract
The advent of wireless networks has accelerated the deployment of large-scale, heterogeneous, and latency-sensitive applications, ranging from Internet of Things infrastructures to real-time cyber-physical systems. While these advances enable pervasive connectivity and intelligence, they also introduce complex security vulnerabilities that traditional centralized solutions cannot address due to privacy concerns, scalability limitations, and susceptibility to single points of failure. This thesis presents a suite of intelligent, privacy-preserving, and attack-resilient federated learning frameworks tailored to the unique security challenges of IoT environments.The research employs a design science methodology, guided by iterative system development, implementation, and empirical evaluation across multiple benchmark datasets, including several IDS datasets, various image datasets, and two fake news corpora. In this study, we propose a robust framework using federated learning for the IDS structure. This novel framework incorporates key artefacts such as a recurrent neural network with a reputation mechanism to mitigate adaptive label-flipping attacks, a graph-based aggregation scheme that utilises client similarity and spectral clustering to defend against poisoning attacks, a federated anomaly detection framework integrating large language models with robust client selection to enable low-latency, high-accuracy detection in wireless network environments, a blockchain-enabled self-supervised FL framework, and a privacy-preserving natural language processing model for misinformation detection.
Comprehensive experiments demonstrate that these frameworks significantly outperform baseline FL methods such as FedAvg, Krum, and Multi-Krum in terms of accuracy and robustness under adversarial conditions. For instance, SNCOC restored up to 98% accuracy after severe poisoning, FedLLMGuard achieved detection accuracies exceeding 99% with under 20 ms mitigation times, BCH-FedSSL maintained stability under label-scarce and adversarial settings, and HFBF consistently outperformed CNN and LSTM baselines in fake news detection tasks. These results collectively validate the feasibility of combining FL with advanced defense mechanisms, blockchain, self-supervision, and LLMs to build resilient, decentralized security solutions.
This thesis contributes to both the theoretical and practical advancement of secure FL. It not only proposes novel architectures for defending against adversarial threats but also establishes trust, transparency, and scalability in decentralized environments. By bridging the gap between robust AI models and privacy-aware deployment, the research lays a foundation for next-generation IDS and misinformation detection frameworks that can operate securely and efficiently in real-world wireless network ecosystems.
| Date of Award | 17 Dec 2025 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Rahim Taheri (Supervisor), Mohammad Shojafar (Supervisor) & Ivan Jordanov (Supervisor) |