Device-Performance-Driven Heterogeneous Multiparty Learning for Arbitrary Images.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Distributed Systems

Background:

• Multiparty learning (MPL) facilitates privacy-preserving collaborative model training.
• Increasing user numbers exacerbate data and model heterogeneity in MPL.
• Existing MPL frameworks struggle with diverse data sizes and device computational capacities.

Purpose of the Study:

• To propose a novel personal multiparty learning method, HMPL, addressing data and model heterogeneity.
• To develop adaptive strategies for unifying feature maps from devices with arbitrary data sizes.
• To enable customized model generation based on individual device performance.

Main Methods:

• Introduced a heterogeneous feature-map integration method for adaptive unification of diverse feature maps.
• Proposed a layer-wise model generation and aggregation strategy for performance-driven customization.
• Implemented aggregation rules for updating shared model parameters based on semantic layer correspondence.

Main Results:

• The HMPL framework effectively addresses both data and model heterogeneity in multiparty learning.
• Experimental results on four datasets demonstrate superior performance compared to state-of-the-art methods.
• The proposed methods show significant improvements in collaborative learning with heterogeneous devices.

Conclusions:

• HMPL offers a robust solution for privacy-preserving collaborative learning in heterogeneous environments.
• The framework's adaptive strategies enhance model performance and generalization.
• This work advances the field of multiparty learning by tackling practical heterogeneity challenges.