Ensembles of Pruned Deep Neural Networks for Accurate and Privacy Preservation in IoT Applications

Alhalabi, Besher (2023) Ensembles of Pruned Deep Neural Networks for Accurate and Privacy Preservation in IoT Applications. Doctoral thesis, Birmingham City University.

Besher Alhalabi PhD Thesis published_Final version_Submitted Jul 2023_Final Award Nov 2023.pdf - Accepted Version

Download (4MB)


The emergence of the AIoT (Artificial Intelligence of Things) represents the powerful convergence of Artificial Intelligence (AI) with the expansive realm of the Internet of Things (IoT). By integrating AI algorithms with the vast network of interconnected IoT devices, we open new doors for intelligent decision-making and edge data analysis, transforming various domains from healthcare and transportation to agriculture and smart cities.

However, this integration raises pivotal questions: How can we ensure deep learning models are aptly compressed and quantised to operate seamlessly on devices constrained by computational resources, without compromising accuracy? How can these models be effectively tailored to cope with the challenges of statistical heterogeneity and the uneven distribution of class labels inherent in IoT applications? Furthermore, in an age where data is a currency, how do we uphold the sanctity of privacy for the sensitive data that IoT devices incessantly generate while also ensuring the unhampered deployment of these advanced deep learning models?

Addressing these intricate challenges forms the crux of this thesis, with its contributions delineated as follows:

Ensyth: A novel approach designed to synthesise pruned ensembles of deep learning models, which not only makes optimal use of limited IoT resources but also ensures a notable boost in predictability. Experimental evidence gathered from CIFAR-10, CIFAR-5, and MNIST-FASHION datasets solidify its merit, especially given its capacity to achieve high predictability.

MicroNets: Venturing into the realms of efficiency, this is a multi-phase pruning pipeline that fuses the principles of weight pruning, channel pruning. Its objective is clear: foster efficient deep ensemble learning, specially crafted for IoT devices. Benchmark tests conducted on CIFAR-10 and CIFAR-100 datasets demonstrate its prowess, highlighting a compression ratio of nearly 92%, with these pruned ensembles surpassing the accuracy metrics set by conventional models.

FedNets: Recognising the challenges of statistical heterogeneity in federated learning and the ever-growing concerns of data privacy, this innovative federated learning framework is introduced. It facilitates edge devices in their collaborative quest to train ensembles of pruned deep neural networks. More than just training, it ensures data privacy remains uncompromised. Evaluations conducted on the Federated CIFAR-100 dataset offer a testament to its efficacy.

In this thesis, substantial contributions have been made to the AIoT application domain. Ensyth, MicroNets, and FedNets collaboratively tackle the challenges of efficiency, accuracy, statistical heterogeneity arising from distributed class labels, and privacy concerns inherent in deploying AI applications on IoT devices. The experimental results underscore the effectiveness of these approaches, paving the way for their practical implementation in real-world scenarios. By offering an integrated solution that satisfies multiple key requirements simultaneously, this research brings us closer to the realisation of effective and privacy-preserved AIoT systems.

Item Type: Thesis (Doctoral)
17 July 2023Submitted
30 November 2023Accepted
Uncontrolled Keywords: Artificial Intelligence, Deep Learning, Pruning, IoT, Federated Learning, Resource management, Compressing Neural Networks, Graph Embeddings
Subjects: CAH11 - computing > CAH11-01 - computing > CAH11-01-05 - artificial intelligence
Divisions: Doctoral Research College > Doctoral Theses Collection
Faculty of Computing, Engineering and the Built Environment > School of Computing and Digital Technology
Depositing User: Jaycie Carter
Date Deposited: 20 Dec 2023 14:17
Last Modified: 20 Dec 2023 14:17
URI: https://www.open-access.bcu.ac.uk/id/eprint/15070

Actions (login required)

View Item View Item


In this section...