SVFLDetector: a decentralized client detection method for Byzantine problem in vertical federated learning

In recent years, with the deepening of cross-industry cooperation, vertical federated learning with multiple overlapping samples and fewer overlapping features has attracted extensive attention. Vertical federated learning increases the challenge of detecting Byzantine clients due to feature heterog...

Full description

Saved in:
Bibliographic Details
Published in:Computing 2024-05, Vol.106 (5), p.1659-1679
Main Authors: Xu, Jiuyun, Jiang, Yinyue, Fan, Hanfei, Wang, Qiqi
Format: Article
Language:English
Subjects:
Artificial Intelligence
Clients
Computer Appl. in Administrative Data Processing
Computer Communication Networks
Computer Science
Federated learning
Heterogeneity
Information Systems Applications (incl.Internet)
Machine learning
Penalty function
Regular Paper
Software Engineering
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In recent years, with the deepening of cross-industry cooperation, vertical federated learning with multiple overlapping samples and fewer overlapping features has attracted extensive attention. Vertical federated learning increases the challenge of detecting Byzantine clients due to feature heterogeneity, in contrast to horizontal federated learning. Existing methods for detecting Byzantine clients can be divided into statistical-based and detection-based types. The detection-based type breaks the limit on the number of Byzantine clients. To our knowledge, current research in vertical federated learning relies on the assumption of a reliable third-party coordinator and is based on statistical type. In this work, we propose a framework based on a detection type called SVFLDetector to detect Byzantine clients in vertical federated learning. The key ideas of SVFLDetector are: (1) we combine decentralized vertical federated learning with split learning, utilizing their respective advantages and eliminating the impact of a third-party server; (2) according to the heterogeneity of features in vertical federated learning, we use a client detection method which is achieved by grouping through feature encoding and performing cross validation within groups to identify Byzantine clients; (3) we propose a penalty function to reduce the impact of Byzantine clients on model aggregation. Numerical experiments show that our method has strong robustness against various Byzantine attacks.
ISSN:0010-485X
1436-5057
DOI:10.1007/s00607-024-01262-5