一种非独立同分布问题下的联邦数据增强算法

Federated Data Augmentation Algorithm for Non-independent and Identical Distributed Data

在联邦学习中,由于用户的本地数据分布会随着用户所在地以及用户偏好而变动,数据的非独立同分布下的用户数据可能缺少某些标签类别的数据,在模型聚合中显著影响了迭代更新速率和最终的模型性能。为了解决这一问题,提出了一种基于条件生成对抗网络进行联邦数据增强的算法,能够在不涉及泄露用户隐私的前提下,通过生成对抗网络模型对数据偏斜的参与者扩增少量数据,大幅提升非独立同分布数据划分下联邦学习算法的性能。实验结果表明,与当前主流的联邦算法相比,该算法在非独立同分布设置下的MNIST,CIFAR-10数据集上的预测精度分别提升了1.18%和14.6%,显示出了该算法对非独立同分布问题的有效性和实用性。

In federated learning,the local data distribution of users changes with the location and preferences of users,the data under the non-independent and identical distributed(Non-IID)data may lack data of some label categories,which significantly affects the update rate and the performance of the global model in federated aggregation.To solve this problem,a federated data augmentation based on conditional generative adversarial network(FDA-cGAN)algorithm is proposed,which can amplify data from participants with skewed data without compromising user privacy,and greatly improve the performance of the algorithm with Non-IID data.Experimental results show that,compared with the current mainstream federated average algorithm,under the Non-IID data setting,the prediction accuracy of MNIST and CIFAR-10 data sets improves by 1.18%and 14.6%,respectively,which demonstrates the effectiveness and practicability of the proposed algorithm for Non-IID data problems in federated learning.