Contribution Measurement in Privacy-Preserving Federated Learning

Federated learning (FL) is a novel decentralized machine learning framework that differs from traditional centralized machine learning. It enables multiple participants to collaborate on training models without sharing raw data directly. Participants train the local model with their data and only upload the model parameters. To achieve a fair allocation of benefits by the common global model, a standard is needed to evaluate each model providers' contribution to federated learning. Shapley value is a classic concept from cooperative game theory and is often used in data evaluation for machine learning. This study introduces the Shapley value in privacy-preserving federated learning (PPFL) to construct a contribution measuring module for measuring the contribution of each model provider to the learning task and propose a verification mechanism for the contribution results. Compared to the other related works for contribution measurement by Shapley value, this work achieves higher privacy protection, where local participants' data sets, local model parameters, and global model parameters are concealed. In addition, the verification of the fairness of contribution measurement is also supported. Moreover, this work achieves the access control of aggregated global models through the concept of threshold identity-based encryption, where model consumers can only gain access to the specific aggregated global model if they are authorized by sufficient model providers.