Research on Optimization of Privacy Protection Algorithm Based on Federated Learning

Faced with the increasing demand and application scenarios for open data sharing in industries such as banking, securities, and insurance, data from different institutions often exhibit high complementarity. However, due to privacy protection and compliance issues, these institutions cannot directly share data. To address this challenge, the article introduces the federated learning optimization algorithm, which enables cross-agency cooperation and allows different institutions to train models without exposing raw data. This approach enhances the effectiveness and accuracy of the models. However, the large-scale implementation of federated learning is hindered by high communication costs, client heterogeneity, low algorithm efficiency, and unstable model performance. Therefore, it is crucial to better apply federated learning algorithms to practical problems and leverage their advantages in data privacy protection and model performance improvement. This article presents a federated learning algorithm specifically designed for optimizing communication costs. Through this optimization algorithm, only one round of communication between the client and server is required. The results demonstrate that, while ensuring accuracy, the amount of communication data is reduced to 1/10 of the Federal Average (FedAvg) algorithm on the dataset, and 1/100 of the FedAvg algorithm on the CIFAR-10 dataset.