SWATM: Contribution-Aware Adaptive Federated Learning Framework Based on Augmented Shapley Values

Abstract

The vanilla federated learning (FedAvg) takes a weighted aggregation of uploaded models based on the amount of data from each participant. By addressing varying data quality and heterogeneous data distributions across participants, the existing work calculates the contribution of participants based on Shapley Value (SV) and attempt to guide model aggregation under SV. Our empirical study found that vanilla SV exhibits low sensitivity and instability issues. To address these issues, we propose a Shapley Value Weighted FL aggregation method based on Adaptive Temperature augmentation and Momentum update (SWATM). The adaptive temperature moderately augments the sensitivity of SV according to aggregation performance, while the momentum update slows down the fluctuation and improves stability. Experiments on CIFAR-10 show that, our approach can significantly improve stability compared to the state-of-the-art, and experiments on household safety inspection show that our method accelerate the convergence of the global model. Finally, we deploy SWATM for predictive maintenance of equipment with a combined F1-score improvement of 2.57%.