Privacy-Preserving Method of Big Data Based on Federated Learning

Abstract:
Big data has brought great convenience to our life, but at the same time, there are also many problems, such as “isolated data island” and “data privacy security”, which can not be ignored. Federated learning can cooperate with multiple participants to complete collaborative training without uploading local data, which can solve these problems well. However, there are still some problems such as insufficient privacy security and low training efficiency. Therefore, this paper designs a Privacy-Preserving Method of Big Data Based on Federated Learning. Firstly, a privacy-preserving protocol based on secret sharing and gradient mask is designed to protect the interactive model parameters in the training process. Secondly, a local training strategy based on SVRG is designed to improve the training efficiency. At last, we prove on MNIST and CIFAR-10 that this method can not only ensure data privacy security without affecting the accuracy of calculation results, but also reduce communication overhead, improve training efficiency and resist gradient leakage attacks.