Federated Learning Scheme Based on Gradient Compression and Local Differential Privacy
Lun Zhao, Siquan Hu, Zhiguo Shi · 2023
As a deep learning framework for distributed deployment, federated learning allows users to complete model training while retaining the original data, which meets the local data privacy protection needs of each user. Differential privacy, as a classic data desensitization technology, is often used to improve privacy protection. Many studies have applied it to machine learning, but it is accompanied by the problem of model utility. In addition, in the traditional federated learning model training, there are frequent gradient exchanges between the user and the central server, which poses a huge challenge to the communication overhead of the model and affects the training efficiency of the model. In order to solve these problems, we propose a new federated learning scheme, firstly propose a StaComp compression algorithm to compress the gradient parameters before the user uploads them, use the local differential privacy mechanism to add noise to protect the user's local data. Finally, we evaluate our scheme on the MNIST dataset. Experiments show that the scheme of this article is close to the classic FedAvg algorithm in the aspect of model accuracy, and is better than other methods based on differential privacy. At the same time, it significantly reduces the training time, improves the convergence speed of the model, and saves communication costs.