TOWARDS PRACTICAL VERTICAL FEDERATED LEARNING

Abstract

The increasing need for high-quality data to effectively train sophisticated machine learning models, particularly when dealing with sensitive and geographically dispersed data, has made federated learning (FL) a vital approach for privacy-conscious environments. Vertical federated learning (VFL), where parties manage distinct features of the same instances, encounters specific obstacles: the absence of extensive benchmarks with varying feature split, performance loss in precise record linkage methods, and elevated communication expenses. To tackle these issues, we developed VertiBench, a benchmarking system that enhances the accuracy of VFL analyses. Additionally, we introduced FedSim, a novel training paradigm that integrates the linkage process directly into the training phase, thereby boosting performance. We also designed FedOnce, a one-shot VFL algorithm designed to reduce communication challenges, and FeT, a scalable model that enables learning from vertically aligned data without the necessity for precise linkage. Collectively, these innovations significantly advance the practical application and development of VFL.