基于联邦学习的玉米叶片病害识别方法

Identification of Maize Leaf Diseases Based on Federated Learning

联邦学习可利用分布式数据实现模型共享训练,无需本地数据上传继而可保证数据资产安全,但数据异构导致本地模型产生漂移而难以有效聚合全局模型。为此,本文提出了一种基于联邦学习的分布式病害识别方法G-FedAvg。针对各用户间数据种类缺失异构导致模型泛化性减弱的问题,通过改进损失函数梯度更新策略,提升用户模型学习捕获全局泛化信息能力;针对数据特征差异导致模型过度拟合,通过自监督预训练,缓解因其所致性能下降。试验以玉米叶片病害识别为导向,并进一步评估病害程度,其结果表明,改进算法G-FedAvg在无需数据上传前提下,取得了与集中学习模型近乎一致的识别性能;与传统联邦学习相比,G-FedAvg的识别准确率与收敛速度有效提升,准确率波动显著降低。因此,所提算法G-FedAvg可有效联合参与用户利用其本地数据完成分布式学习,实现对玉米叶片病害的精准识别。

Federated learning can implement model sharing training by using distributed data,and can guarantee the security of data assets without local data uploading,however,data heterogeneity leads to local model drift and it is difficult to aggregate global models effectively.In this paper,we propose a distributed disease identification method called G-FedAvg.Aiming at the problem that the generalization of the model is weakened due to the lack of heterogeneous data types among users,the loss function gradient updating strategy is improved to improve the ability of learning and capturing global generalization information.In view of overfitting caused by data feature differences,self-supervised pre-training was used to alleviate performance degradation caused by overfitting.The results showed that,the improved algorithm G-FedAvg can achieve almost the same recognition performance without data uploading as the centralized learning model;compared with the traditional federated learning models,G-FedAvg can effectively improve recognition accuracy and convergence speed;meanwhile,the accuracy fluctuation can be significantly reduced.Therefore,the proposed algorithm G-FedAvg can achieve a distributed learning by utilizing the local data of participating users,and the achieved model can realize an accurate recognition of maize leaf diseases.