Feature Map Transfer: Vertical Federated Learning for CNN Models

Federated learning provides a privacy-preserving mechanism for multiple participants to collaboratively train machine learning models without exchanging private data with each other. Existing federated learning algorithms can aggregate CNN models when the dataset is horizontally partitioned, but cannot be applied to vertically partitioned datasets. In this work, we demonstrate the image classification task in the vertical federated learning setting where each participant holds incomplete image pieces of all samples. We propose an approach called VFedConv to solve this problem and achieve the goal of training CNN models without revealing raw data. Different from traditional federated learning algorithms sharing model parameters in each training iteration, VFedConv shares hidden feature maps. Each client creates a local feature extractor and transmits the extracted feature maps to the server. A classifier model at the server-side is constructed with extracted feature maps as input and labels as output. Furthermore, we put forward the model transfer method to improve final performance. Extensive experiments demonstrate that the accuracy of VFedConv is close to the centralized model.