一种基于全同态加密的联邦学习方案

In recent years, federated learning has evolved to provide a new solution for privacy preservation. During the training process, federated learning passes models and gradients rather than sample data as a way to protect the privacy of the sample data. Although federated learning has avoided direct exposure of data, many methods have been proposed in recent years to infer user-sensitive data from the exchanged intermediate parameters. The core idea of this paper is to improve the classical federated learning scheme based on the latest fully homomorphic encryption technology in all aspects to obtain security, functionality and practicality improvements. Based on the functionality of the fully homomorphic encryption algorithm, our federated learning scheme can support high-precision approximation of complex loss functions and can cope with more complex training objectives and tasks; it can significantly improve the training efficiency while better protecting the data security of each participant.