<p>Vertical federated learning (VFL) is a distributed machine learning technique that enables collaborative training between an active party with labeled features and passive parties with unlabeled features. While VFL ensures data privacy and improves model performance, it lacks an effective and efficient mechanism for users to selectively withdraw their contributions from model updates. In this paper, we propose an inverse correction optimization (ICO) vertical federated unlearning scheme, which allows users to selectively remove their data contributions, whether it involves all or only a subset of their data. Specifically, we frame the unlearning problem as an optimization challenge and develop gradient ascent and gradient correction algorithms tailored to the unique training framework of vertical federated learning. Our approach significantly accelerates the retraining speed of the unlearning model while maintaining the model’s effectiveness. Furthermore, we introduce backdoor attacks during training to evaluate the unlearning effectiveness. Experimental results across various tasks and datasets show that ICO achieves a prediction accuracy that differs by only 3.31% from retraining, with an attack performance differing by just 0.49%. While delivering comparable unlearning performance, ICO also accelerates the unlearning optimization process.</p>