Enhancing Federated Learning: Transfer Learning Insights

Federated Learning (FL) is a decentralized machine learning framework that builds a shared model by distributing training data across mobile devices and aggregating updates from local computation. This technology is becoming increasingly popular in distributed medical research collaborations. Although FL performs well in the medical field, medical datasets are usually small, which limits the FL algorithms in terms of performance. Therefore, a challenge associated with datasets is how to select appropriate models to mitigate the possible impact of small datasets. The advantage of transfer learning is that it is borrowing the architecture of a trained model and then fine-tuning certain layers to reduce the computational cost of the training task. However, whether or not transfer learning is more effective than simply using a convolutional neural network model is still a question worth investigating for many applications. In this task, this article proposes an approach that applies a modified transfer learning model on small medical datasets. Theoretically, by utilizing existing robust models, it can capture important features of relevant data to compensate for small datasets. In experiments, this article compares the performance of models using transfer learning with models not using transfer learning in the FL framework. Specifically, this article compared the accuracy of three algorithms under both models, and the results show that the model using transfer learning does not perform as well as the simple convolutional neural network model in terms of performance. So this article explored and analyzed the possible remote causes in depth.