Privacy-Preserving Vertical Federated Logistic Regression without Trusted Third-Party Coordinator
Huizhong Sun, Zhenya Wang, Yuejia Huang, Junda Ye · 2022
Federated learning is a new distributed learning paradigm, which allows multiple parties to cooperatively train a centralized model without sharing their data. In this paper, a privacy-preserving logistic regression (LR) training algorithm for vertical federated learning (VFL) is proposed. First, this paper analyzes the related works and point out the privacy leakage risks. Then, based on the mini-batch SGD and parameter encryption method, a secure VFL model training scheme for LR without the assistance of a trusted third-party is designed. Next, to protect the privacy of model parameters, a differentially-private algorithm and comprehensive privacy analysis are provided. Finally, experiments show that the algorithm not only guarantees the security and privacy, but also ensures the model utility.