A Federated Learning-Based Data Augmentation Method for Privacy Preservation Under Heterogeneous Data

Federated learning is an important distributed machine learning paradigm. This study proposes a privacy-preserving data augmentation model for federated learning of heterogeneous data, which is able to mitigate heterogeneity and augmenting the participant’s local data while protecting data privacy. First, to address the problem of global model bias due to heterogeneous data, this study proposes a distributed generative adversarial network FedEqGAN. The model introduces a multi-source data feature fusion mechanism, which can learn the features of each data source to generate synthetic data. Second, addressing the privacy leakage issue caused by the disclosure of data distribution information, this paper proposes an encryption algorithm for heterogeneous environments FedHE, which utilizes homomorphic encryption to protect local data distributions and aggregates local data information through KL dispersion in order to construct global data distributions. Finally, for the privacy leakage problem caused by uploading model parameters in federation training, this paper proposes a federation model parameter encryption algorithm DPFedMP. This algorithm dynamically injects Gaussian noise into the model parameters according to the difference of data distribution to realize differential privacy protection and update the global model. Experiments show that the method is applicable to heterogeneous data environment, significantly enhancing model performance while ensuring data security.