Graph Neural Networks (GNNs) are essential for handling graph-structured data, often containing sensitive information. It’s vital to maintain a balance between data privacy and usability. To address this, this dissertation introduces three studies aimed at enhancing privacy and utility in GNN applications, particularly in node classification, link prediction, and graph classification. The first work tackles celebrity privacy in social networks. We develop a novel framework using adversarial learning for link-privacy preserved graph embedding, which effectively safeguards sensitive links without compromising the graph’s structure and node attributes. This approach is validated using real social network data. In the second work, we confront challenges in federated graph learning with non-independent and identically distributed (non-IID) data. We introduce PPFL-GNN, a privacy-preserving federated graph neural network framework that mitigates overfitting on the client side and inefficient aggregation on the server side. It leverages local graph data for embeddings and employs embedding alignment techniques for enhanced privacy, addressing the hurdles in federated learning on non-IID graph data. The third work explores Few-Shot graph classification, which aims to classify novel graph types with limited labeled data. We propose a unique framework combining Meta-learning and contrastive learning to better utilize graph structures in molecular and social network datasets. Additionally, we offer benchmark graph datasets with extensive node-attribute dimensions for future research. These studies collectively advance the field of graph-based machine learning by addressing critical issues of data privacy and utility in GNN applications.