SVFL: Secure Vertical Federated Learning on Linear Models

Kaifeng Luo, Zhenfu Cao, Jiachen Shen, Xiaolei Dong

Abstract: Federated learning (FL) is a popular technique that enables multiple parties to train a machine learning model collaboratively without disclosing the raw data to each other. A vertically partitioned federated learning configuration is applicable in a variety of real-world scenarios. In this configuration, a comprehensive feature collection is established only when all parties’ datasets are merged and only one party has access to the labels. Existing vertical federated learning strategies for linear models are not very practical, since they involve either a trusted third-party authority (TPA) or heavy communication overheads. To address this issue, this paper proposes SVFL, a secure vertical federated learning framework on linear models, which is based on the Verifiable Inner-Product Computation (VIP) protocol. SVFL enables the secure and private training of linear models, as well as the validation of a malicious server’s computation. In addition, it decreases the number of communication rounds to 3 and is resistant to collusion attacks. Experiments are done on a variety of real-world datasets from the UCI ML repository, and the results demonstrate that SVFL achieves comparable accuracy to conventional linear models.