Compared with traditional machine learning,federated learning effectively solves the problems of user data privacy and security protection,but a large number of model exchanges between massive nodes and cloud servers will produce high communication costs.Therefore,cloud-edge-side layered federated learning has received more and more attention.In hierarchical federated learning,D2D and opportunity communication can be used for model cooperation training among mobile nodes.Edge server performs local model aggregation,while cloud server performs global model aggregation.In order to improve the convergence rate of the model,the network transmission optimization technique for hierarchical federated learning is studied.This paper introduces the concept and algorithm principle of hierarchical federated learning,summarizes the key challenges that cause network communication overhead,summarizes and analyzes six network transmission optimization methods,such as selecting appropriate nodes,enhancing local computing,reducing the upload number of local model updates,compressing model updates decentralized training and parameter aggregation oriented transimission.Finally,the future research direction is summarized and discussed.

Constructing a novel model for the storage task allocation problem of federated learning client datasets,to ensure load balancing of decentralized cloud storage networks,shorten the storage data uploading and recovery time,and reduce the total client storage cost,a data storage task allocation algorithm——URGL_allo (allocation based on user requirements and global load) that considers client requirements and global load is proposed.In the node allocation phase,node resources such as global load,topological attributes,storage price and data recovery time concerned by clients are considered,and a new node ranking method is defined in conjunction with the law of gravity to select the best storage task allocation node.In the link allocation stage,the shortest path calculation is performed using Dijkstra’s algorithm for the client node as the center to other nodes in the network,and the path with the largest bandwidth value in the set of shortest paths between two nodes is selected for allocation.Simulation results show that the proposed algorithm reduces the load balancing index and the total client storage cost by 41.9% and 5%,respectively,compared with the random policy-based allocation algorithm (Random_allo),and the data recovery time is not much different from that of the link bandwidth-based greedy algorithm,both of which are stably maintained between (0,2],which is 1/20 of Random_allo.The combined performance of global load and service quality is better than that of the comparison algorithm.

Federated learning offers new ideas for solving the problem of multiparty joint modeling in “data silos”.Federated support vector machines can realize cross-device support vector machine modeling without local data,but the existing research has some defects such as insufficient privacy protection in a training process and a lack of research on nonlinear federated support vector machines.To solve the above problems,this paper utilizes the stochastic Fourier feature method and CKKS homomorphic encryption system to propose a nonlinear federated support vector machine training(PPNLFedSVM) algorithm for privacy protection.Firstly,the same Gaussian kernel approximate mapping function is generated locally for each participant based on the random Fourier feature method,and the training data of each participant is explicitly mapped from the low-dimensional space to the high-dimensional space.Secondly,the model parameter security aggregation algorithm based on CKKS cryptography ensures the privacy of model parameters and their contributions during the model aggregation process.Moreover,the parameter aggregation process is optimized and adjusted according to the characteristics of CKKS cryptography to improve the efficiency of the security aggregation algorithm.Security analysis and experimental results show that the PPNLFedSVM algorithm can ensure the privacy of participant model parameters and their contributions to the training process without losing the model accuracy.

In federated learning,the local data distribution of users changes with the location and preferences of users,the data under the non-independent and identical distributed(Non-IID) data may lack data of some label categories,which significantly affects the update rate and the performance of the global model in federated aggregation.To solve this problem,a federated data augmentation based on conditional generative adversarial network(FDA-cGAN) algorithm is proposed,which can amplify data from participants with skewed data without compromising user privacy,and greatly improve the performance of the algorithm with Non-IID data.Experimental results show that,compared with the current mainstream federated average algorithm,under the Non-IID data setting,the prediction accuracy of MNIST and CIFAR-10 data sets improves by 1.18% and 14.6%,respectively,which demonstrates the effectiveness and practicability of the proposed algorithm for Non-IID data problems in federated learning.

Federated learning can effectively ensure the privacy and security of data because it trains data locally on the client.The study of federal learning has made great progress.However,due to the existence of non-independent and identically distributed data,unbalanced data amount and data type,the client will inevitably have problems such as lack of accuracy and low training efficiency when using local data for training.In order to deal with the problem that the federal learning efficiency is reduced due to the difference of the federal learning background,this paper proposes an efficient federated learning scheme based on background optimization to improve the accuracy of the local model in the terminal device,so as to reduce the communication cost and improve the training efficiency of the whole model.Specifically,the first device and the second device are selected according to the diffe-rence in accuracy in different environments,and the irrelevance between the first device model and the global model (hereafter we collectively refer to as the difference value) is taken as the standard difference value.Whether the second device uploads the local model is determined by the value of the difference between the second device and the first device.Experimental results show that compared with the traditional federated learning,the proposed scheme performs better than the federated average algorithm in common federated learning scenarios,and improves the accuracy by about 7.5% in the MINIST data sets.In the CIFAR-10 data set,accuracy improves by about 10%.

Federated Learning(FL) is driven by multi-party data participation,where participants and central servers continuously exchange model parameters rather than directly upload raw data to achieve data sharing and privacy protection.In practical applications,the accuracy of the FL global model relies on multiple stable and high-quality clients participating,but there is an imba-lance in the data quality of participating clients,which can lead to the client being in an unfair position in the training process or not participating in training.Therefore,how to motivate clients to participate in federated learning actively and reliably is the key,which ensuring that FL is widely promoted and applied.This paper mainly introduces the necessity of incentive mechanisms in FL and divides the existing research into incentive mechanisms based on contribution measurement,client selection,payment allocation and multiple sub-problems optimization according to the sub-problems of incentive mechanisms in the FL training process,analyzes and compares existing incentive schemes,and summarizes the challenges in the development of incentive mechanisms on this basis,and explores the future research direction of FL incentive mechanisms.

As a new computing paradigm,edge computing provides computing and storage services at the edge of the network compared to traditional cloud computing model.It has the characteristics of high reliability and low latency.However,there are still some problems in privacy protection and data processing.As a distributed machine learning model,federated learning can well solve the problems of inconsistent data distribution and data privacy in edge computing scenarios,but it still faces challenges in equipment heterogeneity,data heterogeneity and communication,such as model offset,the convergence effect is poor,and the calculation results of some devices are lost.In order to solve the above problems,a federated learning optimization algorithm with dynamic weights(FedDw) is proposed,which focuses on the service quality of the equipment,reduces the heterogeneous impact caused by the participation of some equipments due to inconsistent training speed,and determines the proportion in the final mo-del aggregation according to the service quality,so as to ensure that the aggregation results are more robust in complex real situations.Through experiments,the two excellent federated learning algorithms,FedProx and Scaffold,are compared on the real data sets of 10 regional weather stations.The results show that the FedDw algorithm has better comprehensive performance.

In consideration of the characteristics of multi-dimensional resource consumption fluctuating with time in the Internet of Vehicles system and users’ demands for efficient computing services and data privacy and security,this paper proposes a me-thod of multi-dimensional resource allocation for Internet of Vehicles based on federated learning.On the one hand,the allocation of computing,cache and bandwidth resources is considered comprehensively to ensure the completion rate of computing tasks and avoid the redundant allocation of multidimensional resources.For this purpose,a deep learning algorithm is designed to predict the consumption of various resources through the data collected by edge servers.On the other hand,considering the data island problem caused by users’ data privacy and security requirements,federated learning architecture is adopted to obtain a neural network model with better generalization.The proposed algorithm can not only adjust the allocation of multi-dimensional resources over time,but also meet the resource requirements that change over time,and ensure the efficient completion of computing tasks in the Internet of Vehicles system.Experimental results show that the algorithm has the characteristics of fast convergence and good model generalization,and can complete the aggregation of federated learning with fewer communication rounds.

People commonly hypothesize that an independent variable follows a Gamma distribution in many areas,including hydrology,meteorology and insurance claim.Under the Gamma distribution assumption,Gamma regression model enables an outstanding fitting effect,compared with multivariate linear-regression model.Previous studies may be able to obtain a Gamma regression model trained only on a public dataset.However,when the datasets are provided by multiple parties,how to seek to address the problem of data privacy by training Gamma regression model without exchanging the data itself? A secure multi-party federated Gamma regression algorithm has been applied to this area.Firstly,the log-likelihood function is derived with the iterative method.Secondly,the link function is determined according to the fact,and the gradient updating strategy is constructed by the loss function.Finally,the parameters with homomorphic encryption are updated,then the training is completed.The model is tested on two public datasets,and the results show that under the premise of privacy protection our method can effectively use the value of multi-party data to generate Gamma regression model.The fitting performance of our method is better than that of Gamma regression model implements in a single part,and is close to the result yielded by centralized data learning model.

Automation and accurate fault detection and diagnosis of HVAC systems is one of the most important technologies for reducing time,energy,and financial costs in building performance management.In recent years,data-driven fault detection and diagnosis methods have been heavily studied for fault detection and diagnosis of HVAC systems.However,most existing works deal with single systems and are unable to perform cross-system fault diagnosis.In this paper,a federal learning-based fault detection and diagnosis method is proposed,which uses convolutional neural networks to extract information features,aggregates features using special-designed algorithms,and perform cross-level and cross-system fault detection and diagnosis via federal lear-ning.For multi-fault level fault detection and diagnosis,federal learning is performed using data from four fault levels of chillers.Experimental results show that the average F1-score of the fault detection and diagnosis effect of the four-fault levels is close to 0.97,which is within the practical range.Federal learning uses chiller and air handling unit data for cross-system fault detection and diagnosis.Experimental results show that federal learning using different system data improves the diagnosis results of particular faults,e.g.,14.4% for RefOver faults and 2%~4% for both Refleak and Exoil faults.