Incentive Analysis for Agent Participation in Federated Learning

Lihui Yi, Xiaochun Niu, Ermin Wei

Introduction

Federated Learning (FL) has emerged as a powerful paradigm for collaborative machine learning, allowing multiple agents—such as healthcare institutions or mobile devices—to train a shared model without exposing sensitive local data. While much research has focused on the technical challenges of communication efficiency and robust aggregation, a fundamental economic question remains: why should a self-interested agent participate?

Existing incentive mechanisms often focus on compensating agents for the computational and communication costs incurred during training. However, this paper identifies a more intrinsic factor: data heterogeneity. In FL, the global model is an aggregate of diverse local distributions. If an agent's local data is significantly different from the global average, the resulting global model may perform worse on that agent's specific tasks than a model trained purely on its own local data. This research models the decision-making process of agents as a game-theoretic problem to understand when participation is rational and how decentralized systems reach stability.

The Game-Theoretic Model

The authors define the participation problem through a stage game where each agent $i$ chooses whether to join the federated group or train locally. The decision is governed by a cost function derived from theoretical FL performance bounds. This function balances two competing forces:

  1. The Benefit of Scale: Joining the FL group increases the total amount of data available for training, which generally reduces the model's variance and improves generalization.
  2. The Penalty of Heterogeneity: If the global data distribution deviates significantly from agent $i$'s local distribution, the resulting bias increases the agent's local cost.

Equilibrium and Social Welfare

A central contribution of this work is the characterization of the Nash Equilibrium (NE). The authors prove that agents with similar data distributions naturally form a "consecutive" group in the equilibrium state. Essentially, "birds of a feather flock together"—agents whose data is close to the mean of the participating group find it beneficial to stay, while outliers find it more advantageous to opt out.

Crucially, the paper addresses the "Price of Anarchy"—the gap between selfish behavior and the social optimum. In many systems, individual optimization leads to collective inefficiency. However, the authors demonstrate (Theorem 2) that under mild conditions, the Nash Equilibrium in this FL game perfectly aligns with the strategy that maximizes Social Welfare (minimizing the total cost across all agents). This suggests that a decentralized FL system can be inherently efficient without requiring a central authority to mandate participation.

The Myopic Strategy: Decentralized Convergence

In a real-world decentralized environment, agents lack perfect information about others' data distributions and cost functions. To address this, the authors propose a myopic strategy based on best-reply dynamics.

In this repeated game setting:

  • The server broadcasts only two aggregate values: the number of current participants and their average data mean.
  • Agents use these values to calculate whether they would have achieved a lower cost in the previous round by switching their participation status.
  • Agents update their choices simultaneously.

This approach is notably privacy-preserving, as agents never share their private cost functions or local data distributions. Furthermore, the authors prove that this simple, "bounded rationality" approach converges to a neighborhood of the Nash Equilibrium in finite time. Numerical simulations indicate that convergence is remarkably rapid, often occurring within just one or two stages, regardless of the initial participation density.

Conclusion and Future Directions

This research provides a theoretical foundation for understanding agent incentives in FL through the lens of data quality. By proving that self-interested participation can lead to socially optimal outcomes, the authors offer a positive outlook for decentralized FL ecosystems.

However, the study acknowledges certain limitations. The current proofs rely on a constant data separation parameter ($\Delta$), assuming agents' data means are evenly distributed. Real-world data is rarely so uniform. Future work will focus on relaxing these assumptions to accommodate more complex, non-homogeneous data distributions and generalized payoff structures, bringing these theoretical insights closer to practical, large-scale FL deployments.

Summary updated

BibTeX
@inproceedings{yi2025incentive,
  author = {Lihui Yi and Xiaochun Niu and Ermin Wei},
  title = {{Incentive Analysis for Agent Participation in Federated Learning}},
  booktitle = {CDC},
  year = {2025},
  doi = {10.1109/cdc57313.2025.11312956},
}
Figure