Federated learning (FL) is a distributed machine learning technique in which multiple clients cooperate to train a shared model without exchanging their raw data. However, heterogeneity of data distribution among clients usually leads to poor model inference. In this paper, a prototype-based federated learning framework is proposed, which can achieve better inference performance with only a few changes to the last global iteration of the typical federated learning process. In the last iteration, the server aggregates the prototypes transmitted from distributed clients and then sends them back to local clients for their respective model inferences. Experiments on two baseline datasets show that our proposal can achieve higher accuracy (at least 1%) and relatively efficient communication than two popular baselines under different heterogeneous settings.