MetaMoE: Diversity-Aware Proxy Selection for Privacy-Preserving Mixture-of-Experts Unification

2026-05-13

AI Summary

This work addresses the challenge of integrating independently trained, distributed domain experts into a unified Mixture-of-Experts (MoE) model under strict privacy constraints. To this end, the authors propose MetaMoE, a framework that leverages public proxy data in lieu of private data to unify expert models while preserving privacy. MetaMoE introduces a diversity-aware proxy data selection mechanism that balances domain relevance and distributional coverage to better approximate the underlying private data distribution. Additionally, it incorporates a context-aware router to enable precise expert assignment for heterogeneous inputs. Experimental results demonstrate that MetaMoE significantly outperforms existing privacy-preserving MoE approaches on benchmark datasets in both computer vision and natural language processing.

Abstract

Mixture-of-Experts (MoE) models scale capacity by combining specialized experts, but most existing approaches assume centralized access to training data. In practice, data are distributed across clients and cannot be shared due to privacy constraints, making unified MoE training challenging. We propose MetaMoE, a privacy-preserving framework that unifies independently trained, domain-specialized experts into a single MoE using public proxy data as surrogates for inaccessible private data. Central to MetaMoE is diversity-aware proxy selection, which selects client-domain-relevant and diverse samples from public data to effectively approximate private data distributions and supervise router learning. These proxies are further used to align expert training, improving expert coordination at unification time, while a context-aware router enhances expert selection across heterogeneous inputs. Experiments on computer vision and natural language processing benchmarks demonstrate that MetaMoE consistently outperforms recent privacy-preserving MoE unification methods. Code is available at https://github.com/ws-jiang/MetaMoE.

Problem

• Mixture-of-Experts
• privacy-preserving
• federated learning
• proxy data
• expert unification

Innovation

• Mixture-of-Experts
• privacy-preserving
• proxy data selection
• federated learning
• diversity-aware