Federated learning for next generation intelligent applications

Khan, Ahsan Raza (2025) Federated learning for next generation intelligent applications. PhD thesis, University of Glasgow.

Abstract

The rapid proliferation of smart devices and Internet of Things (IoT) technologies has revolutionised data collection for artificial intelligence (AI)-driven applications, enabling rapid training and near real-time inference. However, the traditional centralised learning approaches require transferring vast amounts of raw data from end devices to a central server. This process leads to substantial network overhead, increased latency, and significant privacy concerns, hindering the scalability and responsiveness of intelligent applications. This thesis exploits federated learning (FL) as a distributed, on-device learning framework that enables collaborative model training without raw data sharing. The distributed architecture of FL offers privacy by design and reduces communication costs by exchanging the model parameters that align with principles of data sovereignty and regulatory compliance. Despite its advantages, FL faces significant challenges in real-world applications, and this thesis aims to address the following three critical challenges: C1) data diversity; C2) robust aggregation ensuring privacy and security in the training process; and finally, C3) energy efficiency. The first contribution introduces the similarity-driven truncated aggregation (SDTA) framework, designed to tackle challenges C1 and C2. SDTA measures the similarity among the model updates to identify and filter the anomalous updates, mitigating the impact of attacks and overfitting without accessing client data. Additionally, it incorporates differential privacy (DP) to strengthen training privacy. The second contribution introduces the semantic-aware federated blockage prediction (SFBP) framework, addressing challenges C1 and C3. Using multi-modal fusion and a lightweight computer vision model for edge-based semantic extraction, the proposed framework reduces communication costs and inference delays while maintaining high prediction accuracy. Additionally, SFBP incorporates a filter mechanism to minimise the effects of noisy or adversarial updates. The third contribution addresses C1 and C3 and develops a hybrid neuromorphic federated learning (HNFL) framework for outdoor human activity recognition (HAR) using wearable sensors. The proposed spiking-long short-term memory (S-LSTM) model combines the energy-efficient spiking neural networks with the sequential data handling strengths of LSTM networks. This approach improves the accuracy while ensuring data privacy and reducing computational costs, making it suitable for deployment on resource-constrained edge devices. Finally, to address challenges C2 and C3, the federated fusion quantisation (FFQ) framework is proposed to improve HAR models in indoor settings. FFQ combines FL with edge-based preprocessing, feature engineering, and model compression to achieve a low false positive rate, essential for applications like fall detection. A customised FedDist algorithm is used for global model aggregation, effectively reducing overfitting in diverse data. Additionally, FFQ applies model compression and quantisation-aware training to lower communication overhead without compromising accuracy. These contributions advance FL by enhancing scalability, robustness, and efficiency, paving the way for next-generation intelligent systems.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Keywords: Federated learning, energy efficiency, robust aggregation, semantic information processing, vision-aided wireless communication, neuromorphic computing.
• Subjects: T Technology > T Technology (General)
• Colleges/Schools: College of Science and Engineering > School of Engineering
• Supervisor's Name: Zoha, Dr. Ahmed, Imran, Professor Muhammad, Mohjazi, Dr. Lina and Flynn, Professor David
• Date of Award: 2025
• Depositing User: Theses Team
• Unique ID: glathesis:2025-85028
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 16 Apr 2025 10:38
• Last Modified: 16 Apr 2025 10:40
• Thesis DOI: 10.5525/gla.thesis.85028
• URI: https://theses.gla.ac.uk/id/eprint/85028

Related URLs

• Article DOI
• Article DOI
• Article DOI
• Article DOI
• Article DOI
• Article DOI
• Article DOI
• Article DOI
• Article DOI
• Article DOI
• Conference proceeding
• Conference proceeding
• Conference proceeding
• Conference proceeding
• Conference proceeding
• Conference proceeding
• Conference proceeding
• Conference proceeding
• Book section
• Book section