Privacy-Preserving Asynchronous Vertical Federated Learning Algorithms for Multiparty Collaborative Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• Federated learning for vertically partitioned (VP) data enables secure multiparty joint modeling without a trusted third party.
• Existing VP federated learning algorithms predominantly use synchronous computation, limiting efficiency with unbalanced resources.
• There is a need for asynchronous VP federated learning algorithms that maintain data privacy and improve performance.

Purpose of the Study:

• To develop and analyze asynchronous federated learning algorithms for VP data.
• To ensure data privacy and provide theoretical convergence guarantees for the proposed algorithms.
• To enhance the efficiency of federated learning in environments with heterogeneous computational and communication resources.

Main Methods:

• Proposed asynchronous federated stochastic gradient descent (AFSGD-VP) algorithm.
• Developed two variance-reduced variants: SVRG and SAGA for VP data.
• Conducted convergence analyses under strong convexity without staleness restrictions.
• Evaluated model privacy, data privacy, computational complexity, and communication costs.

Main Results:

• AFSGD-VP and its SVRG/SAGA variants are the first asynchronous federated learning algorithms for VP data with theoretical guarantees.
• Convergence analyses confirm theoretical performance under strong convexity and without staleness limitations.
• Experimental results on VP datasets validate theoretical findings.
• Proposed algorithms demonstrate significantly higher efficiency compared to synchronous counterparts.

Conclusions:

• The developed asynchronous federated learning algorithms offer a privacy-preserving and efficient solution for VP data.
• These algorithms provide theoretical guarantees, addressing limitations of existing synchronous methods.
• The findings pave the way for more practical and scalable multiparty joint modeling applications.