Evaluate the Contribution of Multiple Participants in Federated Learning

Abstract

To address the challenge of distributed data source, Federated Learning meets with great demand of algorithmic predictions and decisions without taking a risk of privacy leakage, leaving data valuation a consequent task. Establishing an effective profit distribution model enables multiple participants to get involved in a fair incentive. Shapley Value serves as an excellent measure for calculating the contribution of the model since it provides a fair dividend of payoffs. However, it fails under the existence of data replication or dataset partition. In this work, we design a function to recalculate Shapley Value overcoming the failure mentioned before. The testing experiments have proved that new calculation improves the SV performance for about 50% compared with the original index such as the model accuracy or total loss.