Cogo,Vinicius VielmoNeves,Nuno Fuentecilla Maia FerreiraFernandes,Diogo Alexandre Nascimento2026-02-062026-02-062025http://hdl.handle.net/10400.5/116898Tese de mestrado, Engenharia Informática, 2025, Universidade de Lisboa, Faculdade de CiênciasThe proliferation of heterogeneous devices has expanded the attack surface of modern networks, exposing the limitations of traditional Intrusion Detection Systems (IDS). Signature-based approaches, while effective against known threats, struggle to detect novel attacks and depend on constant updates. Machine learning has improved detection, but centralised training raises privacy concerns and remains vulnerable to adversarial interference. Federated Learning (FL) addresses these issues by enabling collaborative training without centralising data, yet it faces challenges of heterogeneity, lack of personalisation, and poisoning attacks. This dissertation proposes a robust, context-aware IDS based on hybrid topology personalised FL. The system combines two phases: an unstructured peer-to-peer phase for decentralisation and scalability, and a structured tree topology where superpeers coordinate aggregation. A reputation mechanism governs superpeer selection, penalising malicious or unreliable peers, while a flooding protocol supports scalable communication. Each peer enhances adaptability by integrating global contributions with personalised local training. Evaluation on the Kitsune dataset assessed performance under non-IID data, personalisation, and adversarial resistance. Results show that personalisation improved detection accuracy across heterogeneous clients, while the reputation mechanism constrained malicious peers and reduced instability. The system consistently achieved around 80% accuracy in multi-class detection tasks, outperforming non-personalised and reputation-absent configurations. These findings demonstrate that combining Federated Learning, personalisation, and reputationbased trust management enhances the robustness, scalability, and adaptability of IDS. The proposed framework provides a resilient and privacy-preserving approach to intrusion detection, suitable for deployment in distributed and adversarial real-world network environments.application/pdfengIntrusion Detection SystemsPeer-to-PeerPersonalised Federated LearningReputation MechanismMachine Learningmaster thesis204177251