Incentives in Federated Learning with Heterogeneous Agents

Federated learning promises significant sample-efficiency gains by pooling data across multiple agents, yet incentive misalignment is an obstacle: each update is costly to the contributor but boosts every participant. We introduce a game-theoretic framework that captures heterogeneous data: an agent’s utility depends on who supplies each sample, not just how many. Agents aim to meet a PAC-style accuracy threshold at minimal personal cost. We show that uncoordinated play yields pathologies: pure equilibria may not exist, and the best equilibrium can be arbitrarily more costly than cooperation. To steer collaboration, we analyze the cost-minimizing contribution vector, prove that computing it is NP-hard, and derive a polynomial-time linear program that achieves a logarithmic approximation. Finally, pairing the LP with a simple pay what you contribute rule, where each agent receives a payment equal to its sample cost, yields a mechanism that is strategy-proof and, within the class of contribution-based transfers, is unique.

💡 Research Summary

Federated learning (FL) enables multiple parties to jointly train a global model while keeping their raw data locally. In realistic settings, however, the data held by each participant are heterogeneous: the utility of a sample depends on which party supplied it, because each party cares primarily about performance on its own distribution. Existing incentive mechanisms for FL largely assume data exchangeability and model utility that depends only on the total number of contributed samples, which fails to capture this heterogeneity.

The paper introduces a game‑theoretic model that explicitly incorporates heterogeneous data. There are k agents, each with a marginal distribution D_i over the instance space and a per‑sample cost c_i. Agent i chooses an integer m_i of labeled examples to contribute. All contributed examples are pooled, and an empirical risk minimizer (ERM) is trained on the pooled set. Each agent i requires that the resulting global model satisfy a Probably Approximately Correct (PAC) accuracy guarantee on its own distribution: with probability at least 1‑δ the error must be ≤ ε. This requirement is encoded as a binary variable a_i(m) that equals 1 if the PAC condition holds and 0 otherwise. The utility of agent i is u_i(m)=a_i(m)−c_i·m_i, i.e., the agent receives a unit payoff if its PAC target is met and pays linearly for each contributed sample.

Key theoretical contributions

1. Equilibrium pathology – The induced non‑cooperative game can lack pure Nash equilibria (Theorem 6). Even when a pure equilibrium exists, its total cost can be arbitrarily larger than the socially optimal cost (the “price of stability” can be infinite). This demonstrates that free‑riding can not only increase individual costs but may also cause the entire federation to become inefficient or even collapse.

2. Social optimum is NP‑hard – Assuming a central planner with full knowledge of all distributions and costs, the problem of finding the cheapest contribution profile that satisfies every agent’s PAC constraint is shown to be NP‑hard (Theorem 2). The hardness stems from the fact that each sample’s source matters, turning the problem into a set‑cover‑like combinatorial optimization.

3. Logarithmic approximation via linear programming – Despite the hardness, the PAC constraints can be linearized, yielding a linear program (LP) that minimizes total cost subject to linearized accuracy constraints. The LP solution is provably within an O(log k) factor of the optimal cost (Theorem 3). The LP is sparse, often assigning most of the contribution burden to a small subset of agents whose data are most valuable to others.

4. Strategy‑proof mechanism – The authors pair the LP allocation with a simple “pay‑what‑you‑contribute” transfer rule: each agent receives a payment equal to its incurred cost c_i·m_i. Under this rule, agents have no incentive to misreport their private data distributions; any deviation reduces expected utility, establishing strategy‑proofness. Moreover, within the class of contribution‑based transfers, this mechanism is uniquely characterized by the combination of strategy‑proofness and cost‑efficiency (Theorem 7).

Empirical validation

Using the FEMNIST benchmark, the authors construct a two‑agent scenario where each agent’s data are biased toward different characters. They empirically confirm that an agent’s test loss decreases monotonically as a larger fraction of the training set originates from its own distribution, illustrating the non‑exchangeability of data. Simulations comparing the LP‑based allocation with naive equal or random contributions show substantial reductions (30‑50 %) in total labeling cost while still meeting every agent’s PAC target.

Discussion and limitations

The analysis assumes an ideal ERM oracle, whereas practical FL systems employ iterative algorithms such as FedAvg that only approximate the pooled ERM solution. The authors discuss how their results extend to these settings and provide additional experimental details in the appendix. The cost model is linear; real‑world costs may include non‑linear privacy risks, communication overhead, or latency penalties. Extending the framework to richer cost structures and multi‑objective settings (e.g., energy consumption) is identified as future work.

Conclusion

The paper delivers the first comprehensive treatment of incentive alignment in federated learning with heterogeneous agents under PAC accuracy guarantees. It reveals severe inefficiencies in uncoordinated play, proves the computational hardness of the optimal cooperative solution, and offers a practical, near‑optimal LP‑based allocation combined with a uniquely strategy‑proof payment rule. These contributions provide both theoretical insight and actionable mechanisms for deploying federated learning systems where participants have diverse data and self‑interested motives.