Abstract: Traditional recommendation systems requires users to share their data, which can potentially lead to privacy concerns. Introducing federated learning into recommendation systems can protect user privacy, but current federated recommendation models require multiple interactions of user-item data to learn global parameters. This paper proposes a federated recommendation model that integrates hypernetworks and item attributes. Firstly, hypernetworks are used to rapidly initialize recommendation models on client devices, with user preferences embedded as input to obtain personalized user preferences. Then, in the client-side recommendation model, content embeddings of item attributes are used as global information to provide additional contextual information. Finally, attention residual blocks are employed to learn the importance of different item attributes. Experiments on the Movielens1M, Hetrec-movielens, and Douban datasets demonstrate that this approach outperforms other models with an average improvement of approximately 1.4% in MAE, 2.0% in RMSE, and 4.1% in accuracy.