Alliance Makes Difference? Maximizing Social Welfare in Cross-Silo Federated Learning

💡 Research Summary

This paper investigates a fundamental social dilemma that arises in cross‑silo federated learning (FL), where a set of organizations jointly train a global model under the coordination of a neutral server. Because the final model is a public good—non‑exclusive and non‑rivalrous—any organization can reap its benefits regardless of its contribution. The authors formalize the interaction among organizations as a public‑goods game, defining each organization’s revenue as a function of the model’s precision improvement and its cost as the sum of computation and communication expenses. The utility of organization i is therefore

 U_i(y) = Φ_i(y) – Ψ_i(y_i),

where y = (y_1,…,y_N) denotes the vector of participation rounds chosen by all N organizations.

Under the realistic assumption that a solitary organization’s local training yields negative net utility, the authors prove that the best‑response of every player is to set y_i = 0, leading to a Nash equilibrium in which no organization participates in any global aggregation. However, when all organizations fully participate (y_i = r for every i), the total social welfare P = Σ_i U_i(y) becomes strictly positive and far exceeds the welfare at the Nash equilibrium. This mismatch between individual rationality and collective optimality constitutes a classic social dilemma.

To resolve the dilemma without resorting to external incentive mechanisms (which incur negotiation and operational costs), the paper introduces the Multi‑player Multi‑action Zero‑Determinant (MMZD) strategy. An MMZD player can enforce a linear relation of the form

 α·U_i + β·P + γ = 0

by appropriately choosing its mixed strategy, irrespective of the strategies of the other players. By selecting parameters (α, β, γ) that set the target value of P, a single organization can unilaterally drive the system’s social welfare to the desired maximum. The authors provide a constructive method for computing these parameters and prove that the MMZD strategy does not require any additional computational or communication resources beyond the standard FL protocol.

Beyond the unilateral case, the paper studies a coalition of several organizations that simultaneously adopt the same MMZD strategy, termed an MMZD Alliance (MMZD A). When multiple players enforce the same linear constraint, the feasible set of social welfare values expands, allowing the alliance to achieve a higher upper bound on P than any single MMZD player could attain alone. Theoretical results (Theorems 2 and 3) delineate the conditions under which the alliance’s collective enforcement yields a strictly larger welfare maximum.

Experimental validation is performed with simulated cross‑silo FL scenarios involving 5–10 organizations, varying the number of global rounds r, computation cost coefficients β_i, communication power ρ_i, and upload time τ_i. The simulations confirm three key observations: (1) the Nash equilibrium yields negative total welfare and poor model precision; (2) a lone MMZD player can raise the welfare to the theoretical optimum, matching the fully cooperative outcome; (3) an MMZD A of two or more players surpasses the single‑player optimum, achieving an even higher total welfare. Importantly, the MMZD and MMZD A strategies incur no extra communication overhead, demonstrating practical applicability.

In summary, the contributions of the paper are: (i) a rigorous game‑theoretic formulation of cross‑silo FL that reveals a social dilemma; (ii) the novel application of MMZD strategies to directly control and maximize social welfare without external incentives; (iii) the extension to MMZD alliances that further improve the welfare bound; and (iv) comprehensive theoretical proofs and empirical results that substantiate the effectiveness and cost‑efficiency of the proposed approaches. This work opens a new avenue for designing incentive‑free mechanisms that ensure cooperative behavior and optimal model performance in federated learning environments.