GameSec 2026 paper reveals federated learning flaws
New research shows Shapley value mechanisms in FL fail on social welfare and fairness under real-world data heterogeneity.
Researchers from the University of Michigan (Lina Alkarmi, Po-Yen Chen, Mingyan Liu) have published an extended version of their paper *Understanding Federated Learning Through the Lens of Mechanism Design: The Role of Data Heterogeneity*, accepted at GameSec 2026. The study critiques existing federated learning (FL) incentive mechanisms, particularly the Shapley value mechanism (M^Shap), which is widely used to motivate data sharing while preventing free-riding.
The paper introduces a new Externality mechanism (M^E) designed to maximize social welfare by aligning individual incentives with collective goals. Unlike M^Shap, which often leads to over-contribution and fails to account for realistic outside options (e.g., standalone training), M^E ensures individual rationality even under data heterogeneity. However, the tradeoff is that M^E sacrifices perfect reciprocity (individual fairness) to achieve social efficiency. Empirical simulations validate these findings, illustrating the fundamental tension between reciprocal fairness and collective optimization in FL systems.
- The Shapley value mechanism (M^Shap) in federated learning (FL) fails to maximize social welfare and lacks individual rationality under data heterogeneity, as shown in a new University of Michigan paper (GameSec 2026).
- The proposed Externality mechanism (M^E) addresses these gaps by maximizing social welfare and ensuring individual rationality, but sacrifices perfect reciprocity (individual fairness) for collective efficiency.
- Empirical simulations confirm the theoretical findings, highlighting a tradeoff between reciprocal fairness and social efficiency in FL incentive mechanisms.
Why It Matters
This research reshapes how AI teams design FL incentive mechanisms, balancing collective benefits with individual fairness in real-world deployments.