Vertical federated learning is a distributed machine learning framework, which is mainly applicable to scenarios where the parties have the same user group, but the feature spaces do not overlap. However, it faces the challenge of high communication complexity during training. Existing methods, such as multiple local updates, asynchronous coordination, or data reduction, aim to reduce this overhead. But they often ignore possible dishonesty, such as the possibility that some parties may exaggerate the importance of the data or falsify gradient updates, which may mislead the global optimization process, weaken the stability and security of the model, and thus affect training results. This paper proposes a core-set selection method based on importance sampling with theoretical guarantee. It considers the honesty of parties in optimizing core-set selection. By reducing data size, it lowers the communication complexity in training. Theoretical analysis demonstrates that the proposed method not only mitigates the negative impact of dishonest users on model training, but also achieves clustering performance comparable to that obtained using the full dataset, while significantly reducing data volume and communication complexity. In addition, experiments show that the proposed core-set method outperforms uniform sampling and achieves results close to those using the full dataset.