Source: https://ieeexplore.ieee.org/document/9859921

most VFL methods assume that at least one party holds the complete set of labels
of all data samples.

VFL requires labels to update model parameters via end-to-end backpropagation.

Federated learning [1, 2, 3] enables multiple parties to collaboratively train a machine learning model with privacy preserving mechanisms.

Conventional federated learning (i.e., horizontal federated learning) assumes that each party has the same set of features but different data.

Recently a more realistic scenario has been investigated in federated learning, where different features are vertically distributed across participants, i.e., vertical federated learning.

In VFL, different parties hold the same set of data, while each party only has a disjoint subset of features.

Gu et al. [4] and Zhang et al. [7] de signed asynchronous parallel architectures to allow only one or partial parties to hold the labels (active parties) and pas-
sive parties transmit their intermediate results to the active parties via efficient tree-structured secure aggregation [15, 8]. However, their works [4, 7] assume that every active party all holds a complete set of labels of all data samples, which is
not realistic in many practical scenarios.