【摘要】 为了提高联邦学习训练中的数据保护能力，提出一种基于分布式差分隐私的联邦学习方法。在分布式过程中，采用安全洗牌模型作为可信任的聚合方，在该框架下每个客户端在其数据上运行DP协议并提供给安全洗牌器，洗牌器随机对报告进行混洗，并将洗牌报告的集合发送到服务器进行最终分析。在该过程中引入Rényi divergence的概念，以便更细致地控制隐私损失。最后在MINIST、Fashion-MINIST数据集上对比实验验证了该方法的可行性，结果表明该方法为联邦学习训练过程提供了有效的隐私保护，同时还维持了良好的模型性能。在MINIST和Fashion-MINIST数据集上分别取得了88%和81.8%的精度，相对于纯本地差分隐私方法提高了10%和6%左右。更多还原

【Abstract】 In order to improve the data protection capability in federated learning training, this paper proposes a federated learning method based on distributed differential privacy.In the distributed process, a secure shuffle model is used as a trusted aggregator.In this framework, each client runs the DP protocol on its data and provides it to the secure shuffler.The shuffler randomly shuffles the reports and sends the collection of shuffled reports to the server for final analysis.In this process, the concept of Rényi divergence is introduced to more finely control privacy loss.Finally, the feasibility of this method is verified by comparative experiments on the MINIST and Fashion-MINIST datasets.The results show that this method provides effective privacy protection for the federated learning training process while maintaining good model performance.It achieves 88% and 81.8% accuracy on the MINIST and Fashion-MINIST datasets, respectively, which are about 10% and 6% higher than that of the pure local differential privacy method.更多还原