Game theory model shows how incentive design shapes data sharing in split federated learning

Split Federated Learning (SFL) combines distributed and split training methods to reduce computational burden and improve model convergence. Yet a critical gap remains: understanding how participants strategically decide whether and how much data to share when pursuing their own interests.

The core problem#

In typical SFL deployments, the model owner must make two interconnected choices. First, they select a cut layer that determines how to split computational work between central servers and participating clients—balancing training load while preserving client privacy. Second, they establish incentive structures to motivate participants to contribute data. But these design decisions don't happen in a vacuum; clients observe the offered incentives and energy costs (from both computation and networking) and independently decide how much data to contribute based on their own utility calculations.

A game-theoretic solution#

Researchers modeled this strategic interplay as a single-leader multi-follower Stackelberg game. In this framework, the SFL model owner acts as a "leader" setting the cut layer and incentives, while clients act as "followers" responding with their data-contribution choices. The team proved that clients' decisions reach a stable equilibrium point—a Nash equilibrium—and analyzed what optimal strategies the leader should adopt.

Key findings and trade-offs#

The Stackelberg equilibrium solution maximizes utility for both participants: clients receive appropriate compensation for their effort, and the model owner achieves high model accuracy while keeping computational and network overhead manageable. The framework also incorporates differential privacy safeguards and establishes criteria for selecting the minimum acceptable cut layer—ensuring clients retain the privacy protection they require.

Why it matters#

This work bridges economics and machine learning by formalizing how incentive design, layer selection, and data contribution decisions interact. Rather than assuming clients will unconditionally participate, the approach recognizes that distributed systems must be designed around rational participant behavior. The equilibrium analysis provides practitioners with a principled method to configure federated systems in real-world settings where participants are self-interested.