With the rapid growth of data in the era of big data, concerns about privacy risks during data exchange have led many institutions to impose strict restrictions on data sharing, resulting in isolated “data silos.” Federated learning has emerged as a key solution to this problem, with vertical federated learning (VFL) being one widely adopted paradigm. However, existing secure protocols in VFL mainly target linear models. They are not directly applicable to complex operations such as ReLU activation, convolution, and pooling, limiting the broader applicability of VFL. This paper proposes a privacy-preserving VFL framework for convolutional neural networks (CNNs) that enables two parties to securely train a CNN model over vertically partitioned data to address this challenge. The framework includes secure protocols for convolution, matrix multiplication, fully connected layers, activation functions, pooling layers, and Argmax, leveraging function secret sharing to reduce online latency and improve efficiency in forward propagation. During back-propagation, the loss is locally computed by the label-holding party, and gradients are efficiently updated. Experimental results demonstrate the effectiveness and efficiency of the proposed framework, offering new insights for collaborative CNN training in VFL settings. To the best of our knowledge, this paper is the first work to construct a convolutional neural network model in VFL.