| With the emergence of 5G and even 6G technologies,the popularity of smart devices has been promoted,which in turn generates a large amount of data at the edge of the network,and using these data for distributed training has gradually become the mainstream in the field of deep learning.However,most of these data are unlabeled,and a small amount of labeled data also shows imbalanced data distribution,and using these data for training will lead to biased model classification.How to make full use of these data while ensuring user privacy has become a major challenge for distributed training of deep learning models.In response to this challenge,academics have proposed a federal semi-supervised learning approach,which combines the ideas of federal learning and semi-supervised learning to improve the classification accuracy and generalization ability of models.However,most of the current research on solving class imbalanced has considered how to solve the data imbalance problem from a single perspective of federal learning or semi-supervised learning,and there is a lack of specific research on combining these two approaches to train deep learning models in a distributed manner in a data imbalance environment.To address the above problems,firstly,this thesis proposes Federated Semi-supervised Learning for Extremely Class Imbalanced(FECI),which alleviates the extreme imbalance caused by the inability of some devices to collect data from certain classes by using unlabeled data for data expansion.extreme imbalance problem caused by the inability of certain classes of data to be collected,thus improving the classification accuracy of the federation learning model training.In FECI,class rebalancing on nodes is achieved by using a self-training based local class rebalancing method,and the extreme class imbalance problem of data is mitigated globally by a global model update selection method based on Kullback-Leibler scatter.Secondly,this thesis proposes Federated Semi-supervised Learning for Class Variable Imbalance(FCVI),which can further solve the class variable imbalance problem caused by the change in the number of classes.FCVI uses the federal gradient monitoring method to monitor the model training parameters and derive the variation of each class.FCVI then uses a category-variable mitigation algorithm to mitigate the effects of category number changes from both local and global perspectives to improve the classification accuracy of the model.Compared with other methods,FCVI does not require additional data information and can quickly and accurately reduce the negative impact of class-variable imbalance on distributed training.Finally,this thesis designs and implements a Federated Semi-supervised Learning Framework for Class Imbalance Scenarios(FFCIS)in a real edge intelligence environment consisting of heterogeneous devices,which is mainly used to support class FFCIS is mainly used to support efficient learning in class imbalance scenarios.It is also tested to verify that FFCIS can significantly mitigate the negative impact of class imbalance on model training in various real class imbalance scenarios. |
