Abstract: The problems of data sharing and privacy protection of Industrial Internet of Things (IIOT) were well solved by multi-party collaborative training of federated learning without exchanging the local private data.But the model offset was mainly caused by partial model renewal when traditional federated learning faced the industrial data of Non Independent and Identically Distributed (Non-IID).The federated dynamic weighted learning was proposed to figure out the problem.The method was divided into two basic steps as partial renewal and global convergence.The federated distance algorithm was used to eliminate the influence of the local model with excessive offset in partial renewal phase;the dynamic weighted algorithm was proposed to assign greater training weights dynamically to the local data that was more beneficial to the overall model in global convergence phrase.The impact of local update to model offset and local model offset to global model were carefully considered in federated dynamic weighted learning method.The experimental results verified that the federated dynamic weighted learning method had good results in the face of industrial data of Non-IID.

Key words: Internet of things, privacy protection, federated learning, non-independent and identically distributed data

摘要： 联邦学习在不交换本地数据的情况下可以完成多方协作训练,很好地解决了工业物联网领域数据隐私保护及共享问题。但是传统的联邦学习在面对非独立同分布的工业数据时,会因为局部模型更新导致模型的偏移。针对上述问题,提出非独立同分布工业大数据下联邦动态加权学习方法,该方法分为局部更新和全局聚合两个阶段。在局部更新阶段,利用联邦距离算法消除偏移程度过大的局部模型的影响;在全局聚合阶段,提出动态加权算法,动态的给对全局模型更有利的局部数据分配更大的训练权重。该方法既考虑了局部更新导致的模型偏移程度问题,又兼顾了偏移局部模型对全局模型的影响。通过实验验证了该方法在面对非独立同分布的工业数据时具有良好的效果。

关键词: 工业物联网, 隐私保护, 联邦学习, 非独立同分布数据