EIVBoost: An efficient and interpretable gradient boosting framework for Vertical Federated Learning

Abstract

The rapid growth of internet data and stringent privacy regulations have posed significant challenges to traditional machine learning methods in multi-party collaborative environments due to data silos. In this context, Vertical Federated Learning (VFL) has emerged as a promising solution. Gradient boosting tree-based VFL schemes, in particular, have gained prominence due to their widespread applicability. However, conventional gradient boosting tree models in VFL settings suffer from notable shortcomings, including high communication overhead from frequent interactions, inefficient utilization of computational resources, limited model interpretability, and privacy leakage risks arising from joint modeling. To address these issues, we propose EIVBoost, a novel gradient boosting tree framework. EIVBoost leverages Function Secret Sharing (FSS) to implement a secure comparison protocol, enabling passive parties without labels to generate pseudo-labels through shared functions and train independently, thereby significantly reducing communication overhead and training time while ensuring privacy. Furthermore, through model simplification and rule extraction, EIVBoost aggregates rules from independent models into a globally interpretable decision tree, enhancing model transparency and inference efficiency. Comprehensive security analyses demonstrate that EIVBoost effectively safeguards data privacy. Extensive experiments on real-world datasets confirm that the framework substantially improves training efficiency without compromising model accuracy, offering a robust, secure, and interpretable solution for VFL.