An Efficient Federated Learning Privacy Protection Algorithm based on Model Quantification

Federated learning, as a distributed machine learning method, enables data sets from different clients to train high-quality models without leaving the local area. However, in the process of federation learning to upload model parameters to the server, privacy leakage is easy to occur. At the same time, with the increase of training rounds, in order to achieve the expected effect, the client frequently uploads data to the server during each training, which will cause certain memory consumption. Therefore, to solve the problem of privacy leakage and excessive memory consumption during client upload, this paper proposes an efficient federated learning privacy protection algorithm (FedEMQ) based on model quantification. The algorithm adds noise before the client uploads the model parameters to the server to protect the privacy of the client. The model parameters are then quantized to convert the floating-point data to a low-precision type to reduce its memory consumption. Finally, the experimental show that the system balances model performance and memory consumption well while protecting client privacy.