Federated Learning (FL) enables multiple parties to train a model without sharing data. However, in heterogeneous scenarios where the data distribution amongst the FL participants is non-independent and identically distributed (non-IID), FL suffers from the data heterogeneity challenge which severely degrades the ability of the global model to converge. To solve this problem, we propose a novel data augmentation strategy, named DPSDA-FL, which can aid in homogenizing the local data present on the client’s side. DPSDA-FL improves the training of the global model by leveraging differentially private synthetic data from foundation models. We obtain promising preliminary results on the CIFAR-10 dataset regarding recall of the global model.