Introduction
In cross-silo Federated Learning (FL), organizations (clients) collaborate to train a global model without sharing raw data. To compensate for computational costs and privacy risks, these clients require incentives. While existing research often focuses on splitting rewards among participants within a single round, the temporal allocation of a fixed total budget across multiple communication rounds remains largely unexplored.
This presents a "Goldilocks" challenge: spending too much per round recruits high-quality clients but exhausts the budget prematurely; spending too little allows for more rounds but results in slow progress and lower final accuracy. Because the relationship between budget, participant count, and model utility is often opaque and dataset-dependent, finding the optimal allocation strategy is difficult.
The BARA Framework
The authors propose BARA (Budget Allocation for Reverse Auction), an online algorithm that dynamically learns the optimal budget strategy during training. BARA acts as a manager for existing incentive mechanisms, determining the ideal number of participants for each round to maximize final model utility. Its architecture is built on three pillars:
- Newton’s Polynomial Interpolation: To overcome the limitation of observing only one outcome per round, BARA uses interpolation to synthesize "artificial" training records, creating a richer dataset for optimization.
- Gaussian Process (GP) Modeling: BARA models the complex, noisy relationship between participant count and model utility using a GP with a composite kernel that prioritizes recent observations.
- Bayesian Optimization (GP-UCB): Using the Upper Confidence Bound algorithm, BARA balances exploring new participant counts with exploiting known high-performing ones.
Experimental Evaluation
BARA was evaluated on four standard benchmarks—MNIST, Fashion-MNIST (FMNIST), CIFAR-10, and CIFAR-100—distributed among 20 clients in a non-IID manner to simulate real-world data heterogeneity.
Key Results
BARA consistently outperformed several baselines, including Even Allocation (EA), Monotonically Increasing/Decreasing (MIA/MDA), and Random Allocation (RA).
Experimental results demonstrate that BARA achieves superior accuracy and efficiency. For instance, on MNIST, BARA reached approximately 90% accuracy, significantly higher than the 87% achieved by the standard Even Allocation strategy. Furthermore, BARA reached these higher accuracy levels using the same total reward budget, proving its cost-effectiveness.
Stability and Convergence
A critical feature of BARA is its ability to identify the optimal "sweet spot" for participant counts. As shown in the figures below, after an initial exploration phase, the algorithm's selection for the number of participants stabilizes across all datasets.
Conclusion and Future Work
BARA demonstrates that treating training duration as a variable dependent on budget allocation can significantly improve FL performance. While highly effective, the current framework assumes a fixed total budget and may face computational bottlenecks in the GP update as communication rounds scale into the thousands. Future research could address dynamic budget environments and more complex client heterogeneity involving varying hardware capabilities and bidding behaviors.
Summary updated