Federated Learning under Non-IID Data : Challenges and Solutions for Robust Model Training
Swetha, Maruvada (2025)
2025
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025121235812
https://urn.fi/URN:NBN:fi:amk-2025121235812
Tiivistelmä
Federated Learning (FL) is a decentralized machine learning framework that enables multiple devices to collaboratively train a shared model while keeping their local data private. By avoiding the direct transfer of sensitive data to a central server, FL addresses key concerns related to privacy protection, data ownership, and communication efficiency. However, unlike conventional machine learning environments, federated systems rarely satisfy the assumption of independent and identically distributed (IID) data. In real-world deployments, client data typically exhibit significant variations in size, feature distribution, and class representation, resulting in non-IID conditions.
This statistical heterogeneity introduces several challenges, including unstable training behaviour, biased global models, slow convergence, and limited generalization. As a result, effectively managing non-IID data is a critical requirement for the reliable and large-scale deployment of federated learning.
The main objective of this research is to investigate and develop methods that enhance the robustness of federated learning in the presence of non-IID data. Specifically, this study aims to:
• Examine the impact of non-IID data on convergence and prediction performance using widely adopted benchmark datasets.
• Design and implement improved federated learning techniques, including regularization-based methods such as FedProx, personalized federated learning strategies, clustering-based aggregation methods, and transfer learning approaches.
• Evaluate these methods through experimental analysis on datasets such as MNIST, CIFAR-10, FEMNIST, and selected healthcare datasets, supported by performance visualizations.
• Demonstrate the practical usefulness of the proposed solutions in application areas such as mobile text prediction, medical diagnostics, IoT systems, and financial services, where data heterogeneity is unavoidable.
The findings of this study indicate that well-designed adaptations to the standard federated learning process can significantly reduce the negative effects of non-IID data. Clustering clients with similar data distributions enhances stability and fairness, while personalized learning improves model performance at the individual client level. In addition, transfer learning offers an effective mechanism for adapting pretrained models to highly imbalanced local datasets.
This thesis contributes to ongoing research in federated learning by presenting a comprehensive framework for addressing non-IID challenges through theoretical analysis, algorithmic design, experimental validation, and domain-specific case studies. Beyond its academic contributions, this work supports the development of secure, fair, and scalable federated learning systems for real-world, privacy-sensitive applications in healthcare, mobile computing, IoT, and financial services.
This statistical heterogeneity introduces several challenges, including unstable training behaviour, biased global models, slow convergence, and limited generalization. As a result, effectively managing non-IID data is a critical requirement for the reliable and large-scale deployment of federated learning.
The main objective of this research is to investigate and develop methods that enhance the robustness of federated learning in the presence of non-IID data. Specifically, this study aims to:
• Examine the impact of non-IID data on convergence and prediction performance using widely adopted benchmark datasets.
• Design and implement improved federated learning techniques, including regularization-based methods such as FedProx, personalized federated learning strategies, clustering-based aggregation methods, and transfer learning approaches.
• Evaluate these methods through experimental analysis on datasets such as MNIST, CIFAR-10, FEMNIST, and selected healthcare datasets, supported by performance visualizations.
• Demonstrate the practical usefulness of the proposed solutions in application areas such as mobile text prediction, medical diagnostics, IoT systems, and financial services, where data heterogeneity is unavoidable.
The findings of this study indicate that well-designed adaptations to the standard federated learning process can significantly reduce the negative effects of non-IID data. Clustering clients with similar data distributions enhances stability and fairness, while personalized learning improves model performance at the individual client level. In addition, transfer learning offers an effective mechanism for adapting pretrained models to highly imbalanced local datasets.
This thesis contributes to ongoing research in federated learning by presenting a comprehensive framework for addressing non-IID challenges through theoretical analysis, algorithmic design, experimental validation, and domain-specific case studies. Beyond its academic contributions, this work supports the development of secure, fair, and scalable federated learning systems for real-world, privacy-sensitive applications in healthcare, mobile computing, IoT, and financial services.