FLavourite: Horizontal and Vertical Federated Learning Setting Comparison

Federated learning (FL) enables multiple organizations to collaboratively train AI models while keeping data separate. Depending on data distribution, FL can be set up horizontally (HFL) or vertically (VFL). When multi-modal data are already available in the same infrastructure, the choice of FL setting depends on the workload needs, such as level of privacy, evaluation metrics, and time-to-train. In this paper, we perform a qualitative and extensive evaluation of the HFL and VFL settings to provide guidance on how to choose a specific setting based on workload needs. We assess the classification of Alzheimer’s disease using non-independent and identically distributed (non-i.i.d.) 3D images and tabular data. The results demonstrate how, with the same aggregation function, HFL converges faster to higher accuracy and f1-score than VFL and demonstrate how HFL can be further improved due to a more flexible architecture. Finally, we provide practical considerations on when to choose which setting and share the implementations of the pipelines for reproducibility.