Federated learning (FL) has emerged as a promising paradigm for training machine learning models across distributed devices while preserving data privacy. However, realworld deployment of FL faces several critical challenges, including non-IID data distributions, noisy and mislabelled datasets, resource heterogeneity across participating devices, and class imbalance. These issues substantially degrade model performance, convergence stability, and scalability, limiting the practical applicability of FL in sensitive domains such as healthcare, finance, and education. This thesis addresses these challenges through the development of interconnected frameworks that systematically improve the robustness, adaptability, and inclusivity of federated learning systems. The first major contribution is DQFed (Data Quality-driven Federated Learning), a novel quality-driven aggregation strategy that assigns weights to local model updates based on the statistical quality of client datasets, measured using Shannon Class Entropy. Extensive experiments on CIFAR-10-derived datasets demonstrate that DQFed consistently improves model accuracy under both IID and non-IID conditions, achieving F1-scores of 0.83 and 0.70 respectively, consistently outperforming established baselines such as FedAvg, FedProx, and FedOpt. Building on DQFed, the thesis extends its noise robustness by integrating a Variational Autoencoder (VAE)-based noise detection mechanism and a Noise Reduction Penalty metric. This extension enables DQFed to handle both class imbalance and label noise simultaneously, maintaining high accuracy across heterogeneous and imperfect data environments. Empirical validation across experimental configurations involving up to 25 participating clients, using CIFAR-10-derived data partitions confirms DQFed’s consistently higher performance and model-agnostic applicability across diverse architectures. To address resource heterogeneity, the thesis introduces AH-Fed (Adaptive Heterogeneous Federated Learning), which employs a tri-tiered architecture and stratified resource management. AH-Fed enables devices with varying computational capacities to participate effectively in FL, reducing communication costs by approximately 85% in accuracy-comparable configurations High-resource clients (HRC tier), and up to 99% in the most resource-constrained setting Low-resource clients (LRC tier), albeit with corresponding reduction in model accuracy. This framework substantially improves device inclusivity and system scalability. Finally, the thesis demonstrates DQFed’s domain-specific applicability through an entropybased extension for student dropout prediction in educational analytics. This application addresses strict privacy requirements and class imbalance inherent in educational datasets, achieving improved F1-scores and competitive aggregation times compared to baseline FL strategies. Collectively, these contributions advance both the theoretical foundations and practical implementations of federated learning, enabling more reliable, scalable, and privacypreserving distributed machine learning systems across diverse real-world applications.

Enhancing Federated Learning for AIoT: Robust Aggregation for Imbalanced and Noisy Data in Smart Edge Device Environments / Usman, M.. - (2026 Jul 29).

Enhancing Federated Learning for AIoT: Robust Aggregation for Imbalanced and Noisy Data in Smart Edge Device Environments

Usman Muhammad
2026-07-29

Abstract

Federated learning (FL) has emerged as a promising paradigm for training machine learning models across distributed devices while preserving data privacy. However, realworld deployment of FL faces several critical challenges, including non-IID data distributions, noisy and mislabelled datasets, resource heterogeneity across participating devices, and class imbalance. These issues substantially degrade model performance, convergence stability, and scalability, limiting the practical applicability of FL in sensitive domains such as healthcare, finance, and education. This thesis addresses these challenges through the development of interconnected frameworks that systematically improve the robustness, adaptability, and inclusivity of federated learning systems. The first major contribution is DQFed (Data Quality-driven Federated Learning), a novel quality-driven aggregation strategy that assigns weights to local model updates based on the statistical quality of client datasets, measured using Shannon Class Entropy. Extensive experiments on CIFAR-10-derived datasets demonstrate that DQFed consistently improves model accuracy under both IID and non-IID conditions, achieving F1-scores of 0.83 and 0.70 respectively, consistently outperforming established baselines such as FedAvg, FedProx, and FedOpt. Building on DQFed, the thesis extends its noise robustness by integrating a Variational Autoencoder (VAE)-based noise detection mechanism and a Noise Reduction Penalty metric. This extension enables DQFed to handle both class imbalance and label noise simultaneously, maintaining high accuracy across heterogeneous and imperfect data environments. Empirical validation across experimental configurations involving up to 25 participating clients, using CIFAR-10-derived data partitions confirms DQFed’s consistently higher performance and model-agnostic applicability across diverse architectures. To address resource heterogeneity, the thesis introduces AH-Fed (Adaptive Heterogeneous Federated Learning), which employs a tri-tiered architecture and stratified resource management. AH-Fed enables devices with varying computational capacities to participate effectively in FL, reducing communication costs by approximately 85% in accuracy-comparable configurations High-resource clients (HRC tier), and up to 99% in the most resource-constrained setting Low-resource clients (LRC tier), albeit with corresponding reduction in model accuracy. This framework substantially improves device inclusivity and system scalability. Finally, the thesis demonstrates DQFed’s domain-specific applicability through an entropybased extension for student dropout prediction in educational analytics. This application addresses strict privacy requirements and class imbalance inherent in educational datasets, achieving improved F1-scores and competitive aggregation times compared to baseline FL strategies. Collectively, these contributions advance both the theoretical foundations and practical implementations of federated learning, enabling more reliable, scalable, and privacypreserving distributed machine learning systems across diverse real-world applications.
29-lug-2026
38
Dottorato di Ricerca in Tecnologie dell'informazione per l'Ingegneria
Federated Learning, Non-IID Data, Data Quality, Resource Heterogeneity, Edge Computing, AIoT (Artificial Intelligence of Things), Class Imbalance, Noise Robustness.
BERNARDI, Mario Luca
File in questo prodotto:
Non ci sono file associati a questo prodotto.

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12070/77466
 Attenzione

Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo

Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact