Robust Federated Learning Client Selection Using Shapley Value in Heterogeneous Environments

As a distributed learning approach, Federated Learning (FL) aims at training a federated model while preserving the data privacy of participants. Despite the promising applications of FL, prevailing research grapples with the intricate task of appraising and delineating the weight of each participant’s contribution. In this paper, we propose a novel FL approach called R-FedSV, which leverages Shapley values while incorporating the concept of replay buffers to measure each participant’s contribution to the FL aggregation phase. By calculating the Shapley value (SV) for each participant, we can objectively assess its contribution to the global model, overcoming the limitations of naive averaging used in traditional FL methods. Additionally, Replay Buffer makes our contribution assessment more accurate. We improve accuracy by treating FL training as a cooperative game, adjusting weights by calculating client contributions via shapley values, and storing historical data samples of well-performing clients for re-selection via a replay buffer. Finally, a more generalized global model can be achieved. To validate the efficacy of R-FedSV, we have conducted abundant experiments in real-world non-i.i.d data distribution scenarios. Experimental results show that our method outperforms traditional approaches in terms of model accuracy, and stability.