Federated learning (FL) produces an effective global model by aggregating multiple client weights trained on their private data. However, it is common that the data are not independently and identically distributed (non-IID) across different clients, which greatly degrades the performance of the global model. We observe that existing FL approaches mostly ignore the distribution information of client-side private data. Actually, the distribution information is a kind of structured knowledge about the data itself, and it also represents the mutual clustering relations of data examples. In this work, we propose a novel approach, namely Federated Distribution Knowledge Transfer (FedDKT), that alleviates heterogeneous FL by extracting and transferring the distribution knowledge from diverse data. Specifically, the server learns a lightweight generator to generate data and broadcasts it to the sampled clients, FedDKT decouples the feature representations of the generated data and transfers t