随着人工智能的发展，其成果广泛应用在人们的日常生活中。训练机器学习模型需要使用大量数据，然而数据分布在不同用户之间，其中常包含有 隐私敏感数据。为了在保护数据隐私和安全的同时利用多方数据共同训练模 型，基于差分隐私的联邦学习算法产生。本文围绕降维联邦平均算法展开，该 算法通过提取客户端加噪模型的低维子空间在保护隐私的同时提高联邦优化 效用。本文介绍了算法的实现原理与流程，通过对比实验在联邦学习场景下 验测试算法的准确率，通过投影维度选择实验验证算法提取低维子空间的正 确性，满足了用户在保护数据隐私的同时获得机器学习模型的需求。

With the development of artificial intelligence, its achievements are widely used in People’s Daily life. Training machine learning models requires large amounts of data, which are distributed among different users and often contain privacy-sensitive data. In order to protect data privacy and security while using multi-party data to train model, federated learning algorithm based on differential privacy is developed. This paper focuses on the projected federated averaging algorithm, which can improve the effectiveness of federated optimization while protecting privacy by extracting the low-dimensional subspace from the client model update. In this paper, the implementation principle and process of the algorithm are introduced. The accuracy of the test algorithm is tested in the federated learning scene by comparison experiment, and the correctness of the algorithm to extract low-dimensional subspace is verified by the projection dimension selection experiment, which meets the needs of users to obtain machine learning models while protecting data privacy.