Label Inference Attacks Against Vertical Federated Learning

As the initial variant of federated learning (FL), horizontal federated learning (HFL) applies to the situations where datasets share the same feature space but differ in the sample space, e.g., the collaboration between two regional banks, while trending vertical federated learning (VFL) deals with the cases where datasets share the same sample space but differ in the feature space, e.g., the collaboration between a bank and an e-commerce platform. Although various attacks have been proposed to evaluate the privacy risks of HFL, yet, few studies, if not none, have explored that for VFL. Considering that the typical application scenario of VFL is that a few participants (usually two) collaboratively train a machine learning (ML) model with features distributed among them but labels owned by only one of them, protecting the privacy of the labels owned by one participant should be a fundamental guarantee provided by VFL, as the labels might be highly sensitive, e.g., whether a person has a certain kind of disease. However, we discover that the bottom model structure and the gradient update mechanism of VFL can be exploited by a malicious participant to gain the power to infer the privately owned labels. Worse still, by abusing the bottom model, he/she can even infer labels beyond the training dataset. Based on our findings, we propose a set of novel label inference attacks against VFL. Our experiments show that the proposed attacks achieve an outstanding performance. We further share our insights and discuss possible defenses. Our research can shed light on the hidden privacy risks of VFL and pave the way for new research directions towards more secure VFL.