Research on Federated Learning‐Based Privacy Protection Strategies Under Edge Computing Scenarios

With the advent of billions of end devices connected to the edge of the internet, a large amount of valuable edge data is generated in production and life. When raw data is trained centrally, the risk of privacy leakage exists. Federated learning protects local data privacy by uniting multiple computing nodes for efficient machine learning without sharing data. However, federated learning faces the problem of data sparsity in practical applications, which makes global models difficult to train. In addition, explosive data growth reduces the availability of training models and tends to bring network transmission pressure. For solving these issues, this paper proposed a privacy‐preserving strategy based on joint learning in edge computing scenarios. First, the concept of parallel over‐parameterization is utilized to dynamically sparsely train joint learning models to reduce the resources occupied by communication and computation. Second, the privacy preservation of model parameters is investigated by adding perturbations to model parameters to ensure the security of model uploading to edge servers. Eventually, this paper conducted comparison experiments with HFL, DP‐FedAvg, and SAFL. The experimental results demonstrate that the method proposed in this paper has certain advantages in privacy protection.