Social Welfare Maximization for Federated Learning with Network Effects

TL;DR

This paper introduces a novel incentive mechanism for federated learning that accounts for network effects among clients, significantly improving social welfare through strategic model trading and sharing.

Contribution

It develops a theoretical model of network effects in FL and proposes the SEMTS mechanism to maximize social welfare without extra incentive costs.

Findings

Up to 148.86% improvement in social welfare in experiments

Network effects in FL are non-monotonic and impactful

The SEMTS mechanism effectively aligns incentives and enhances welfare

Abstract

A proper mechanism design can help federated learning (FL) to achieve good social welfare by coordinating self-interested clients through the learning process. However, existing mechanisms neglect the network effects of client participation, leading to suboptimal incentives and social welfare. This paper addresses this gap by exploring network effects in FL incentive mechanism design. We establish a theoretical model to analyze FL model performance and quantify the impact of network effects on heterogeneous client participation. Our analysis reveals the non-monotonic nature of FL network effects. To leverage such effects, we propose a model trading and sharing (MTS) framework that allows clients to obtain FL models through participation or purchase. To tackle heterogeneous clients' strategic behaviors, we further design a socially efficient model trading and sharing (SEMTS) mechanism. Our mechanism achieves social welfare maximization solely through customer payments, without additional incentive costs. Experimental results on an FL hardware prototype demonstrate up to 148.86% improvement in social welfare compared to existing mechanisms.