Research On Key Technologies Of Data Sharing Based On Federated Learning In The Internet Of Things

With the booming development of Io T(Internet of Things,Io T)technology,massive amounts of Io T terminal devices accessing the network have generated large-scale local operational data.However,most of these data are exclusively owned by the holders.On the one hand,it causes the problem of data islands,on the other hand,it hinders the training of high-quality models based on big data and reduces the effect of data service applications.At the same time,the distributed model inference feature of federated learning fits well with Io T,so the technology of Io T data sharing based on the federated learning framework has become a hot research topic.However,in the existing research on the federated learning data-sharing framework,less attention has been paid to the influence of Io T device data and device behavior on the shared models,which easily leads to the problems of low model accuracy and poor security.To address the above problems,this thesis investigates the key techniques for efficient and secure sharing of Io T-oriented federated learning.The specific research work is as follows.(1)This paper studies a layered sharing architecture based on blockchain and federated learning,which considers the communication efficiency of the federated learning model training process and designs a device layer,aggregation layer and computation layer.Among them,the device layer mainly performs local model computation;the aggregation layer selects a central node for local model forwarding based on the idea of clustering,considering communication distance and other constraints,and verifies the block content in the blockchain consensus process;the computation layer is mainly responsible for computing the global model and operating the consensus algorithm process,and finally generates a block containing the global model by consensus.Through the design of device grouping and model consensus,the above architecture achieves efficient and secure Io T data sharing.(2)Aiming at the problem of low accuracy of federated learning model due to the unbalanced distribution of Io T device data labels,this paper proposes a federated learning algorithm for device clustering based on label similarity.First,the method uses the distance between the label distribution of the multi-device combination and the global label distribution to design the label similarity,and the larger the label similarity,the more similar the two distributions are;then performs client node clustering based on the label similarity;further,a sequential training model is designed for the client devices in the cluster.As the updated model parameters of the devices can be shared within the cluster,thus improving the local model accuracy of the model.Finally,the proposed federal learning algorithm for device clustering is validated using the MNIST dataset,and the learning performance and the stability of the learning algorithm are analyzed.The experimental results show that the model accuracy of this paper’s federated learning algorithm can reach 95% when the device data label distribution is in extreme imbalance,i.e.only one class of labels is available.(3)To address the problem of poor security of federated learning models due to unreliable behavior of Io T devices,this paper proposes a trusted consensus algorithm for federated learning model security to prevent unreliable devices from providing false model parameters,which may cause damage to federated learning model security.First,a multi-weighted subjective logical reputation scheme is designed to comprehensively assess the reputation of model interaction nodes to identify malicious nodes and reduce their impact on the global model of federated learning.Then,the node reputation is used to design a trusted consensus algorithm,and the consensus difficulty can be adjusted during the execution of the algorithm to improve transaction efficiency and reduce the time overhead of block generation.The experimental results show that the trusted consensus algorithm in this paper can improve the model accuracy by 6.3% when the probability of maliciousness of a node is 60% compared to the traditional Po W consensus algorithm.