Multiparty Dual Learning.

Area of Science:

• Machine Learning
• Distributed Systems
• Data Privacy

Background:

• Machine learning performance depends on large datasets, often unavailable due to data policies in distributed settings.
• Isolated parties may have insufficient or poor-quality data for effective model training.

Purpose of the Study:

• To propose a novel multiparty dual learning (MPDL) framework to overcome data limitations in isolated parties.
• To enhance machine learning model performance in distributed environments with data constraints.

Main Methods:

• Developed a multiparty dual learning (MPDL) framework leveraging dual task correlations for regularization.
• Introduced feature-oriented differential privacy with mathematical proof to protect raw features during dual inference.
• Designed the framework for minimal modification to existing multiparty learning structures.

Main Results:

• MPDL framework significantly improves performance compared to state-of-the-art multiparty learning methods.
• Achieved accuracy comparable to non-distributed self-learning approaches for isolated parties.
• Demonstrated effectiveness through simulations on real-world datasets.

Conclusions:

• MPDL effectively alleviates the problem of limited and poor-quality data in isolated parties within a distributed learning setting.
• The framework offers a privacy-preserving and flexible solution for multiparty machine learning.
• Dual learning principles are well-suited for addressing missing data challenges in distributed scenarios.