Federated Open Set Domain Adaptation Based On Client Difference

Open set domain adaptation learning which is an emerging research field of transfer learning,aims to transfer knowledge from the source domain with rich annotation data to the target domain with insufficient annotation data and to complete the classification task of known classes in the target domain and the identification of new class samples.During training model,open set domain adaptation learning methods need to use source data and target data at the same time.Generally,source and target data are stored separately.Considering the limitations of data security and privacy protection,it becomes difficult to gather the data together.As a new model training method,federated learning does not require the original data to be sent out of the local client in the training process.Clients can combine all client training models only through the exchange of model related information.Therefore,the concept of federated open set domain adaptation is proposed in this thesis,and two federated open set domain adaptation methods are designed based on client differences.The main research work and contributions of this thesis are as follows:(1)A federated open set domain adaptation method based on client contributions is proposed in this thesis.This method designs the model from the perspectives of local model updating and federated aggregation.In the local model updating stage,the source domain client and the target domain client jointly train the local model by sharing model parameters,and use the adversarial network to train a pseudo boundary between the known class samples and unknown class samples in the target domain.Trained generator can select part of the samples as the known class samples to align with the source domain samples,and select part of the samples as the unknown class samples.In the federated aggregation stage,a federation aggregation strategy based on the contribution of each source domain client and a federated aggregation strategy based on the difference of category information in each source domain client are proposed.The first strategy measures the contribution of each client to the task of the target domain by the gap statistics gain between the client model and the characteristics of the target domain,and assigns weights to the client model according to the contribution in the aggregation process,the second strategy preserves the category information of each client in the process of global model aggregation,and gives full play to the client’s ability to identify specific category samples.(2)A federated open set domain adaptation method based on client uncertainty is proposed in this thesis.In the local model updating stage,the evidence theory is introduced,and the Dirichlet distribution is used to fit the class probability distribution to output the uncertainty of the client.The uncertainty of the client represents the quality of the client.The lower the uncertainty of the client,the more reliable the client is,and the higher the weight allocated in the federated aggregation process.A series of comparative experiments are designed to verify two proposed methods.The experimental results are better than the existing methods.The results fully prove the importance of paying attention to client differences in federated open set domain adaptation learning.