MMVFL: A Simple Vertical Federated Learning Framework for Multi-Class Multi-Participant Scenarios

Area of Science:

• Machine Learning
• Data Privacy
• Distributed Computing

Background:

• Federated learning (FL) enables collaborative model training while preserving data privacy.
• Vertical federated learning (VFL) addresses scenarios with shared samples but distinct features, where one party holds labels.
• Existing VFL research primarily focuses on two-party, binary-class problems, with recent work emphasizing communication and security.

Purpose of the Study:

• To propose a novel framework, multi-participant multi-class vertical federated learning (MMVFL), for multi-class VFL problems with multiple participants.
• To enable privacy-preserving label sharing from the label owner to other participants in a VFL setting.
• To demonstrate the effectiveness of MMVFL, particularly when integrated with feature selection, for multi-class classification tasks.

Main Methods:

• The MMVFL framework extends multi-view learning (MVL) principles to facilitate secure label sharing among multiple VFL participants.
• A feature selection scheme is integrated within MMVFL to evaluate its performance and quantify feature importance.
• The framework allows for the measurement of individual participant contributions to the collective model.

Main Results:

• MMVFL effectively shares label information across multiple VFL participants in a privacy-preserving manner.
• The integrated feature selection within MMVFL demonstrates comparable multi-class classification performance to existing approaches.
• The framework successfully quantifies feature importance and participant contributions, validating its utility.

Conclusions:

• The proposed MMVFL framework is a viable solution for multi-class vertical federated learning involving multiple parties.
• MMVFL facilitates effective label sharing and achieves competitive classification performance while maintaining privacy.
• The framework's modular design allows for easy integration with advanced communication and security techniques.