Robust Decentralized VFL Over Dynamic Device Environment

Surojit Ganguli, Zeyu Zhou, Christopher Brinton, David I. Inouye

Keywords: Robustness, Vertical Federated Learning

Abstract:

Robust collaborative learning on a network of edge devices, for vertically split datasets, is challenging because edge devices may fail due to environment conditions or events such as extreme weather. The current Vertical Federated learning (VFL) approaches assume a centralized learning setup or assume the active party or server cannot fail. To address these limitations, we first formalize the problem of VFL under dynamic network conditions such as faults (named DN-VFL). Then, we develop a novel DN-VFL method called **M**ultiple **A**ggregation with **G**ossip Rounds and **S**imulated Faults (MAGS) that synthesizes faults via dropout, replication, and
gossiping to improve robustness significantly over baselines. We also theoretically analyze our proposed approaches to explain why they enhance robustness. Extensive empirical results validate that MAGS is robust across a range of fault rates—including extreme fault rates—compared to prior VFL approaches.

Primary Area: other topics in machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8544