As a promising distributed machine learning paradigm, Federated Learning brings huge privacy-preserving potentials, and has become a hot topic of research in recent years. To tackle client drift induced by statistically heterogeneous user data, this paper first presents an intermediate feature generation method based on Generative Adversarial Networks for the aim of classifier correction. Secondly, to deal with the particular problem of concept shift, a personalized model aggregation approach is proposed on the basis of classifier clustering. Finally, the two strategies mentioned above are integrated and tested on the CIFAR-10 image classification dataset. Various empirical results show that the proposed integrated strategy, compared to the classic Federated Averaging algorithm, helps realize both better generalization of the single-center global model, and better personalization of the multi-center cluster models.