NTP-VFL - A New Scheme for Non-3rd Party Vertical Federated Learning

Di Zhao, Ming Yao, Wanwan Wang, Hao He, Xin Jin

Abstract: Vertical Federated Learning (FL) handles decentralized and partitioned vertically data about common entities. While most existing privacy-preserving federated learning algorithms require a third party (TP) as an intermediary data accessor to coordinate model training, we propose a new private-preserving scheme named NTP-VFL (Non-3rd Party Vertical Federated Learning). Utilizing Paillier homomorphic encryption, our algorithm strategy allows for multi-party model training and guarantees clients’ privacy against honest-but-curious adversaries. To the best of our knowledge, this is the first non- TP method that solves multi-party computation problems in Logistic Regression tasks. Our theoretical analysis and extensive experiments show outstanding performance with an average increase in efficiency of about 25% baselines with the traditional federated learning approach.