Workshop: 13th Annual International Workshop on Innovating the Network for Data Intensive Science — INDIS
Authors: Evangelos Pipilikas, Alexandros Koufakis, and Ana Oprescu (Universiteit van Amsterdam)
Abstract: Modern scientific collaborations rely on combining data from multiple organizations while preserving data privacy according to regulations and policies. Vertical Federated Learning (VFL) enables collaborative model training without requiring participants to share raw data. However, deploying VFL across geographically distributed infrastructure introduces significant execution overhead, and requires balancing computational cost and privacy requirements.
To address the communication overhead introduced by distributed VFL execution, we investigate the deployment of multiple VFL execution strategies within DYNAMOS, a policy-aware middleware for privacy-preserving data exchange. We evaluate the baseline VFL protocol alongside two communication-efficient variants, including a novel adaptation of an asynchronous optimization technique from Horizontal Federated Learning. The strategies are deployed in both a local environment and on a distributed research testbed, enabling a comparative evaluation under realistic scientific networking conditions. These experiments support execution time, model utility, and privacy analysis, demonstrating that the optimized strategies substantially reduce runtime while preserving predictive performance.
Back to 13th Annual International Workshop on Innovating the Network for Data Intensive Science — INDIS Archive Listing
Back to Full Workshop Archive Listing