A Federated Learning Framework Based on CSP Homomorphic Encryption
Ran Zeng, Bo Mi, Darong Huang
Abstract
In the era of big data, the field of deep learning is developing rapidly. Deep learning model algorithms and a large amount of data make deep learning an effective tool to solve practical problems. However, in the current centralized deep learning model of a large number of client-server models, when we upload our data for training on the server side, there is a risk of leaking privacy. Federated Learning is a type of distributed machine learning that allows multiple institutions or individuals to learn collaboratively without exchanging data. Users only upload their model parameters, and the server aggregates the model parameters of each user into a global model and returns it to the client, and the client updates its local model according to the global model to achieve the global optimal solution, avoiding leakage of private data. However, attackers can still restore user data by obtaining uploaded model parameters from users, which leads to the fact that only transmitting model parameters cannot protect user privacy. Therefore, the issue of privacy protection has become the focus of federated learning. In this paper, the CSP (conjugate search problem) fully homomorphic encryption algorithm is used to encrypt the user model parameters. The CSP encrypt algorithm will not cause the loss of model performance, and at the same time has the characteristics of fast encryption speed, which is very suitable for application in the field of machine learning.