Graph-based vertical federated learning (GVFL) enables collaboration by incorporating scattered attributes and adjacency relations from aligned nodes, and allows each party to contribute its personalized input embedding to joint training and inference. The injection of adversarial inputs can mislead the inference towards attacker’s will, forcing other benign parties to make negligible contributions and losing rewards regarding the importance of their contributions. However, most attacks require server model architectures, queries, or labeled auxiliary graphs from the training domain. These extra requirements are not practical for real-world GVFL applications. In this paper, we propose SGAC, a novel attack framework for crafting adversarial inputs to dominate joint inference without relying on such above requirements. SGAC advances prior attacks by requiring only access to auxiliary graphs from non-training domains. SGAC learns generalized label-indicative embeddings and estimates class-transferable probabilities across domains to generate a surrogate model that closely approximates the server model. SGAC then emphasizes salient node attributes and edges in the auxiliary graph, creating a diverse shadow input set that resembles influential test inputs. With surrogate fidelity and input diversity, SGAC crafts transferable adversarial inputs. Evaluation on diverse model architectures confirms the effectiveness of SGAC.