Social networks have increasingly become essential for communication and community building, resulting in substantial structured graph data. Analyzing such data is heavily dependent on understanding graph structures. Graph embedding techniques can effectively transform complex graph data into low-dimensional vectors, supporting various downstream tasks. However, traditional embedding methods typically neglect privacy, making node representations vulnerable to attribute inference attacks that expose sensitive user information. To address this, adversarial graph autoencoders have incorporated privacy-preserving mechanisms, although they generally rely on centralized training manner, creating potential privacy risks. In this paper, we introduce a novel federated learning framework that integrates adversarial graph autoencoders (FL-AdvGNN), allowing distributed and privacy-preserving generation of node embeddings. We comprehensively evaluate the framework’s performance under realistic, non-IID graph structures and various federated learning architectures. Our empirical results demonstrate that FL-AdvGNN significantly improves privacy protection in federated environments with minimal impact on utility. Furthermore, we address existing research gaps by offering extensive experimental validation and insights into privacy-utility trade-offs in federated adversarial graph learning contexts, providing valuable guidance for future developments in privacy-sensitive applications.