Research And Application Of Semi-supervised Federated Learning Under Non-independent And Identically Distributed Dat

Federated learning(FL)allows edge devices to keep data locally while simultaneously training a shared high-quality global model.However,current research is generally based on an assumption that the training data of local clients have ground-truth.To this end,this paper proposes a semi-supervised federated learning(SSFL)framework that aims to address the data availability problem in FL.Specifically,this paper proposes a model parameter mixing strategy to improve the naive combination of FL and semi-supervised learning methods under two realistic scenarios(labels-at-client and labels-at-server),which is referred to as federated mixing(Fed Mix).Furthermore,in order to improve the performance of the global model in SSFL,this paper proposes an aggregation rule based on the frequency of client participation in training,which can dynamically adjust the weights of the corresponding local models according to the frequency of client training,which is called federated frequency(Fed Freq).With the gradual popularization of self-driving,it is becoming increasingly important for vehicles to smartly make the right driving decisions and autonomously obey traffic rules by correctly recognizing traffic signs.In particular,this emerging technique relies on ground-truth labeled data shared among connected vehicles to train a high-quality automatic image recognition model.However,this approach poses data privacy and data availability challenges.In addition,the local data collected by vehicles using different camera modules in different environments has significant differences in quantity,characteristics,and distribution,namely the non-independent identically distributed(non-IID)problem.To address these issues,this paper proposes a robust semi-supervised federated learning algorithm for privacy-preserving traffic sign recognition,called Fed MB.Specifically,this paper proposes a new consistency regularization loss that mixes strong and weak augmentations to improve the generalization ability of the model.Based on Bayesian estimation,this paper proposes a pseudo-label loss with prior probability knowledge to improve the performance of the model.Furthermore,Fed MB adopts a model mixing strategy to complete the learning of unsupervised model,supervised model and global model to alleviate the non-IID problem in traffic sign image recognition.In particular,this paper studies a challenging traffic sign recognition scenario where labeled and unlabeled data are separated at different locations(i.e.,between the road side unit and the vehicle).Extensive experimental evaluations on two public datasets CIFAR-10 and Fashion-MNIST and a real-world Belgium traffic sign dataset demonstrate that the proposed algorithm outperforms mainstream baselines in performance.In particular,the model recognition accuracy on the Belgium traffic sign dataset with only 3% labeled data reaches 76.8%,and the method has good and stable model performance under different levels of non-IID distribution.