A Study Of Image Classification Algorithm Based On Heterogeneous Federated Learning

In recent years,artificial intelligence technology has been widely applied in many fields such as object detection,image recognition and natural language processing.However,traditional AI techniques have encountered challenges in the isolation of data and personal privacy.To address these issues,federated learning has been proposed.In this thesis,we proposed an image classification algorithm based on a federated heterogeneous architecture that enables each client to design different models according to the distribution of data,thus promoting high-performance federated learning in heterogeneous scenarios.The details as follows:(1)This thesis proposed FedMMD,a federated heterogeneous architecture with multiteacher and multi-feature distillation.Firstly,a federated heterogeneous architecture based on knowledge distillation is proposed to solve the data heterogeneity.In the aggregation phase of federated learning,knowledge distillation allows each client to transmit information extracted from local data to other clients,relaxing the requirement of the federated averaging algorithm that all client models be consistent.Next,the concept of multiteachers is introduced to mitigate the impact of system heterogeneity.By designating a constant number of teachers to each client in every communication based on the principle of similarity,the efficiency of knowledge transfer is improved,and malicious attacks from certain clients can be prevented.Finally,a knowledge distillation algorithm based on multi-feature labels is proposed,under the assumption that a network model can be divided into a feature extractor and a classifier.This algorithm enables the student model to learn not only the soft labels but also the feature information from different semantic depths of the teacher model,thereby extending the dimension of knowledge extracted from the teacher model and enabling the student model to acquire the classification and feature extraction abilities of the teacher model.Experiments conducted on MNIST,EMNIST,CIFAR10,and CIFAR100 show that FedMMD achieves a 1% ～ 5% improvement in accuracy over the federated averaging algorithm.Finally,the FedMMD is migrated to the graph network and the above data set is converted into graph data for experiments.The results showed that the FedMMD algorithm had an accuracy rate 1% ～ 4% higher than the federated averaging algorithm,which proved the universality of the method proposed in this thesis.(2)This thesis proposed a novel federated heterogeneous architecture with DTViT(Visual Transformer with Distillation Token),based on transfer learning.For the purpose of improving the image classification performance of FedMMD,this method combines the traditional visual self-attention model,named ViT and knowledge distillation based federated heterogeneous architectures.Specifically,DTViT uses dual tokens to separate the image classification abilities from different sources.The classification token is used to learn the classification information from own data set and public data set.While the distillation token is used to extract the classification information from the teacher model,which improves the generalization ability of DTViT.Additionally,transfer learning is introduced by loading pre-trained ViT model parameters for the initial of DTViT which can accelerate model convergence.Experiment results show that the proposed DTViT federated heterogeneous architecture with transfer learning outperforms other homogeneous or heterogeneous federated learning algorithms by 1% ～ 15% in the image classification task.