Artificial intelligence (AI) has driven rapid advances in data-driven technologies, but the increasing scale and sensitivity of data raise significant privacy concerns. Distributed Service Framework have emerged to address these challenges, enabling privacy-preserving model training services across multiple sources without centralizing raw data. Among them, federated learning (FL) has gained prominence as a naturally distributed paradigm, allowing collaborative optimization while keeping data localized. However, despite its advantages, FL still faces risks of privacy leakage through gradients and shared updates. To strengthen protection, differential privacy (DP) offers formal guarantees by limiting the contribution of individual records, making it a widely studied solution for secure Distributed services. In parallel, graph neural networks (GNNs) have demonstrated strong capabilities in modeling graph-structured data, though their interdependent structures complicate the direct application of DP and often lead to performance degradation. To address this challenge, we present DP-GNN-P, a framework that partitions the original graph into multiple disjoint subgraphs and injects noise into the gradient updates of each partition. This strategy limits gradient interference among nodes, enabling a more favorable balance between privacy protection and model utility. Experiments on Cora, Citeseer, and Pubmed show that DP-GNN-P maintains strong classification accuracy under diverse privacy budgets. Its partition-based design mitigates noise impact and adapts well to varying graph structures, making it suitable for privacy-sensitive graph learning tasks.