Abstract
We present Poisson Binomial Mechanism Vertical Federated Learning (PBM-VFL),a communication-efficient Vertical Federated Learning algorithm withDifferential Privacy guarantees. PBM-VFL combines Secure Multi-PartyComputation with the recently introduced Poisson Binomial Mechanism to protectparties' private datasets during model training. We define the novel concept offeature privacy and analyze end-to-end feature and sample privacy of ouralgorithm. We compare sample privacy loss in VFL with privacy loss in HFL. Wealso provide the first theoretical characterization of the relationship betweenprivacy budget, convergence error, and communication cost indifferentially-private VFL. Finally, we empirically show that our modelperforms well with high levels of privacy.