Securing intelligent networks: federated learning approaches for privacy-conscious anomaly detection

Manzoor, Habib Ullah (2025) Securing intelligent networks: federated learning approaches for privacy-conscious anomaly detection. PhD thesis, University of Glasgow.

Abstract

As intelligent systems increasingly rely on distributed data generated at the edge, traditional centralized machine learning paradigms face critical limitations related to data privacy, band-width constraints, and vulnerability to single points of failure. Federated Learning (FL) has emerged as a promising distributed alternative that enables collaborative model training across devices without sharing raw data, preserving privacy while leveraging collective intelligence. In smart energy systems, Short-Term Load Forecasting (STLF) exemplifies a critical application where FL’s privacy-preserving capabilities offer significant advantages, supporting efficient grid operation. However, deploying FL in dynamic and heterogeneous environments poses unique challenges, including security vulnerabilities from untrusted participants, communication bottlenecks in low-bandwidth networks, and managing data heterogeneity across clients. This research systematically addresses these hurdles by targeting four key challenges to advance the practical deployment of FL in energy-centric applications. The first challenge (C1) involves model attacks, where adversarial clients attempt to compromise the global model. Existing attacks do not fully exploit FL’s vulnerabilities and can be captured with current defence frameworks. This necessitates the development of stealth attack strategies. This research introduces novel stealth model poisoning techniques, including the Federated Communication Round Attack (Fed-CRA), which increases communication rounds without degrading model performance but at the cost of higher resource consumption. These vulnerabilities highlight the need for stronger defense mechanisms, driving the development of more robust frameworks. The second challenge (C2) focuses on robust aggregation, as traditional FL methods often struggle to filter out adversarial updates effectively. To address this challenge, we introduce four innovative frameworks: (a). Federated Random Layer Aggregation (FedRLA), which enhances security by aggregating only a randomly selected layer during each round, which reduces the attack surface as attack can only attack single layer of local model thereby mitigating the impact of adversarial updates; (b). Layer-Based Anomaly Aware Federated Averaging (LBAAFedAvg), which detects and isolates compromised layers while ensuring that valid updates are preserved with the help novel clustering criteria to identify good and back clients, improving the overall integrity of the aggregation process; (c). Federated Incentivized Averaging (Fed-InA), specifically designed for Fed-CRA, which is based on game theory, it incentivizes honest clients by rewarding them and penalizes malicious ones, promoting a healthier collaborative environment; and Decentralized Federated Learning (DFL), which distributes the aggregation process across multiple clients, minimizing the risk of single points of failure and eliminate the need of server. Furthermore, (d). Decentralized Federated Random Layer Aggregation (DRLA) combines DFL with FedRLA to significantly enhance robustness against adversarial attacks by aggregating a single layer in peer to peer communication manner. The third challenge (C3) concerns communication and computational efficiency, as FL’s iterative updates can strain bandwidth and processing resources, especially in energy-constrained environments. The proposed frameworks optimize efficiency by minimizing transmitted data and computational overhead. FedRLA significantly reduces communication costs by limiting shared model information, while Adaptive Single Layer Aggregation (ASLA) leverages quantization and adaptive stopping criteria to ensure minimal resource usage. Other robust frameworks, LBAA-FedAvg, Fed-InA, and DFL, are designed to require minimal resources for model training. The fourth challenge (C4) addresses data heterogeneity, a fundamental issue in energy networks where clients possess diverse consumption patterns. Two frameworks tackle this problem: (a). FedBranched, which clusters clients based on data similarity to enhance local model convergence, and (b). ASLA, which selectively aggregates the most effective layer across clients, improving generalization across varied datasets. Through addressing these interconnected challenges, this research enhances the robustness of the model, communication and computational efficiency and fixes the issue of data heterogeneity between different clients, paving the way for more resilient and privacy-preserving intelligent systems.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Keywords: Federated learning, energy efficiency, robust aggregation, anomaly detection, data heterogeneity, communication efficiency.
• Subjects: T Technology > T Technology (General)
• Colleges/Schools: College of Science and Engineering > School of Engineering
• Supervisor's Name: Zoha, Dr. Ahmed
• Date of Award: 2025
• Depositing User: Theses Team
• Unique ID: glathesis:2025-85412
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 25 Aug 2025 14:48
• Last Modified: 25 Aug 2025 14:55
• Thesis DOI: 10.5525/gla.thesis.85412
• URI: https://theses.gla.ac.uk/id/eprint/85412

Related URLs

• Article DOI
• Article DOI
• Article DOI
• Article DOI
• Article DOI
• Conference item