Addressing data heterogeneity, scarcity, and training efficiency in privacy-preserving federated learning
Year: 2025
Type: dissertation
Abstract: Over the past decade, deep learning has achieved significant breakthroughs in various domains, including computer vision and natural language processing, with applications spanning industries such as healthcare and manufacturing. However, the success of deep neural networks (DNNs) relies heavily on access to large-scale datasets, which can be difficult for individual organizations to acquire. The reasons behind are the complexities and costs for data collection and annotation at large scale. A straightforward solution is to centralize data from multiple organizations for training. However, this approach raises significant privacy concerns, as such data often contains sensitive or confidential information. Furthermore, regulations like the General Data Protection Regulation (GDPR) emphasize the importance of protecting user privacy in inter-organizational data exchanges. Moreover, transmitting large volumes of data introduces substantial computational overhead, further complicating the process. These challenges highlight the urgent need for methods that facilitate collaborative data use while ensuring privacy preservation. Federated learning (FL), which enables multiple parties to collaboratively train a DNN with the assistance of a central server, offers an effective solution to the aforementioned problem. Unlike traditional centralized learning, which requires collecting data from each party, FL eliminates the need to upload data for joint training. Instead, locally trained models are exchanged with a central server, which aggregates the knowledge from all of the uploaded models and then distributes a refined global model to each party. This approach allows each party to benefit from the collective contributions, ultimately enhancing the model performance. This thesis makes four key contributions to addressing challenges in federated learning, including data heterogeneity, data scarcity, and system convergence rate, while preserving client data privacy. For each contribution, we propose a novel method and empirically demonstrate its effectiveness within the relevant problem context. In this thesis, we address key challenges in federated learning (FL) related to data heterogeneity, foundation model adaptation, hyperparameter tuning, and communication efficiency. First, we tackle feature space heterogeneity by introducing a generative augmentation method that aligns diverse client distributions, validated on both public and real-world datasets. We then adapt parameter-efficient fine-tuning techniques for vision-language models in FL by designing a dual-stream adapter that captures both client-specific and client-agnostic knowledge. To optimize performance under limited resources and non-IID conditions, we propose an evolutionary hyperparameter tuning framework that enables efficient online optimization. Lastly, we address data scarcity and heterogeneity in One-Shot FL by personalizing pretrained latent diffusion models, enabling privacy-preserving synthetic data generation that improves performance across challenging domains such as medical and satellite imaging.
DOI: 10.5282/edoc.36097
Author: Chen, Haokun
Institution: Ludwig-Maximilians-Universität München
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7108654185