Abstract: FedGame is a multiplayer simulation platform designed to study human behavior in response to various Federated Learning (FL) incentive schemes. By simulating a competitive virtual marketplace, the platform enables researchers to gather behavioral data and design reward mechanisms that encourage high-quality data sharing while addressing the economic challenges of FL participation.
Introduction: The Economic Hurdle of Federated Learning
Federated Learning (FL) has emerged as a powerful paradigm for training machine learning models collaboratively while preserving data privacy. However, despite its technical advantages, FL faces a significant economic challenge: providing sufficient incentives for participation. In a competitive marketplace, businesses often view their data as a primary asset. Contributing to a shared FL model may inadvertently benefit rivals, leading to "market share erosion." Furthermore, participants must contend with communication costs and the "free-rider" problem, where low-contributing members reap the benefits of high-quality data provided by others.
To sustain a healthy FL ecosystem, robust incentive schemes are required. Yet, designing these schemes is difficult because human behavior—specifically how participants decide to join, leave, or contribute data under different reward models—is not well understood.
The FedGame Framework
FedGame is a multiplayer simulation platform designed to bridge the gap between theoretical game theory and actual human decision-making in FL environments. It creates a virtual marketplace where "Federations" (groups of players) compete to build the best FL models and capture market share.
The game involves a mix of human participants and AI agents. Human players role-play as business owners with specific data assets and resources. The gameplay proceeds in three distinct stages:
- Bidding: Players evaluate available federations and decide which to join, pledging specific quantities and qualities of data.
- FL Training: The system simulates the collaborative training process based on the participants' contributions.
- Profit-Sharing: The resulting market profits are distributed among federation members according to a pre-defined incentive scheme.
Incentive Mechanisms and Behavioral Data
FedGame allows researchers to test various payoff-sharing models, including:
- Linear: Rewards proportional to data usefulness.
- Equal: Profits split evenly among all members.
- Individual: Payoffs based on marginal contribution to total profit.
- Union: Based on the "Labour Union" game-theoretic model.
- Shapley Value: A fair revenue-sharing approach based on each player's unique contribution to all possible sub-groups.
As a research tool, FedGame's primary output is behavioral data. It records every decision and resource allocation, allowing researchers to model human "rationality" and identify which schemes best encourage high-quality data sharing while minimizing "cheating" or attrition.
Strategic Dynamics and Visualization
The platform is designed to simulate the "Competitor's Dilemma." Players must constantly weigh the benefits of a superior AI model against the risk of empowering their competitors. To ensure realistic decision-making, the system provides a comprehensive real-time dashboard.
This visualization tool displays market share, profit/loss trends, and the quality of other participants. By presenting information similar to what a real-world executive would see, the game captures how information transparency influences strategic behavior.
Conclusion and Future Directions
FedGame demonstrates a successful integration of human decision-making with multi-agent systems to study FL economics. Its extensible architecture, configured via XML, allows for the easy addition of new incentive schemes and environmental variables like communication costs.
While the current version focuses on Horizontal Federated Learning (HFL), future iterations aim to incorporate Vertical Federated Learning (VFL) and Federated Transfer Learning (FTL). Additionally, the authors plan to introduce more complex business parameters, such as legal compliance risks and varying technical overheads, to further enhance the realism of the simulation.
Summary updated