Privacy Preserving on Updated Parameters in Federated Learning
Wenqiang Yang, Bin Liu, Changlei Lu, Nenghai Yu
Abstract
Federated learning provides a framework in which many participants join together to train a deep learning model. Although data is not directly transmitted in federated learning in order to protect privacy, recent researches show that transmitted parameters also lead to information leakage, which violates participants' privacy. In this paper, we combine cryptographic tools (additively homomorphic encryption, AES and RSA) with federated learning to design privacy-preserving protocols, which protect every participant's parameters' information. Results of experiments show that the proposed cryptographic methods can protect single participant's uploaded parameters with acceptable computation overhead increasing.