Adaptive and Efficient Participant Selection in Vertical Federated Learning

Vertical Federated Learning (VFL) enables multiple data owners to collaboratively train a global model, while preserving data privacy. However, significant communication overhead arises during the training phase due to the transmission of massive encrypted intermediate results. Our investigation into the performance of the model with varying participant numbers reveals that increasing participants, especially in larger groups, may not necessarily yield substantial performance improvements. Instead, it tends to introduce additional communication overhead. Therefore, in the context of VFL, participant selection emerges as a crucial strategy to mitigate communication cost, though it remains a relatively unexplored domain. In this work, we propose a novel adaptive participant selection method in VFL, named VFLMG, to dynamically determine the number of selected participants, as well as to identify the group of participants that best strikes a balance between model performance and communication overhead. Evaluations on various datasets show that VFLMG significantly reduces communication overhead while only causing a minimal loss in accuracy. For instance, in the MNIST and Fashion MNIST datasets, VFLMG reduces communication costs by 32.1 % and 27.6 % respectively, with an accuracy drop of 0.5 % and 0.3 %.