PFed-ME: Personalized Federated Learning Based on Model Enhancement

In recent years, with the increasing concern for data privacy- preserving, traditional machine learning methods have encountered numerous challenges, which has prompted federated learning to become the focus of research. Federated learning has the potential to train robust models while safeguarding data privacy and security. However, the performance of current federated learning methods tends to degrade when encountering non-iid situations. This paper proposes a new personalized federated learning algorithm called PFed-ME. The algorithm adaptively chooses to perform model enhancement. Classification technology is used to classify local clients, and generative adversarial network technology is used to optimize the classified model on the server. It aims to alleviate the problem of non-iid data among clients, keep local clients personalized, and finally improve the training accuracy of clients. We evaluate our approach in an image classification task for MNIST and CIFAR10. The results show that our method performs well in terms of accuracy and convergence, outperforming the most advanced methods. In addition, we also verify the effectiveness of this approach on the medical Colon dataset.