Research On Federated Learning Optimization Methods In Edge Computing Scenario

With the popularity of Artificial Intelligence(AI)technology,various intelligent applications and services are flourishing.However,they also bring new challenges.On one hand,the current deployment form of intelligent applications based on centralized cloud computing requires end devices to upload data to cloud data centers for processing,and it not only takes up a lot of network bandwidth but also has slow data transfer and high latency.On the other hand,the form of intelligent application deployed in the cloud with high privacy leakage risk has been increasingly questioned as governments pay more and more attention to users’ data privacy.In this situation,Edge Computing(EC)has emerged in the public’s field of vision,which can leverage the large amount of data in the edge network to train high-quality AI models by collaborating multiple end devices and edge servers for periodic interactions,local training and global aggregation.Federated Learning(FL),a machine learning paradigm that enables data security and privacy protection,has favorably fueled the development of EC with its advantages such as supporting diversity and heterogeneity.However,current end devices are heterogeneous in terms of computation,storage,network transmission and data,leading to shortcomings in the performance deficiencies of the current federated learning methods.At the same time,end devices also face various security threats such as poisoning and escape during the model training process.Therefore,how to design high-performance and secure FL optimization methods in edge computing scenario in the presence of heterogeneous resources,data,and other aspects,and in the face of multiple security threats is an urgent research problem.The paper focuses on the performance optimization of FL methods in edge computing scenario in heterogeneous environments and security enhancement against attacks under the existing foundation and conditions,and the following outlines the primary research contents and innovative work:(1)Research on a D2D-oriented federated learning method for learning efficiency optimizationThe current mainstream FL model training methods directly use the communication topology as a fixed collaborative topology for training,and end devices with poor channel conditions or limited computing power can significantly reduce the efficiency of FL.To solve the problem,a Device-to-Deviceoriented(D2D-oriented)federated learning method for learning efficiency optimization is proposed,which is called D2D-oriented Federated Learning(D2DFL).Firstly,the topology is transformed according to the connectivity of the communication topology.Secondly,in order to maximize the learning efficiency,that is,to maximize the decay of the global loss function in the shortest possible time,the learning efficiency optimization model is proposed,taking into account the end devices’own capabilities and the allocation of the system Resource Block(RB)resources and communication power.To solve the model,a model growth method based on heuristic algorithm is proposed to find the end devices combination model with the maximum learning efficiency as the collaborative topology.Finally,the model is trained according to the determined topology.The experimental results show that D2D-FL can guarantee the accuracy of FL while achieving better results in terms of learning efficiency.(2)Research on a personalized federated learning method based on graph clusteringTraditional FL is all collaborative to obtain a unique global model,which cannot well meet the personalization needs of end devices with a large amount of non-Independent and Identically Distribution(non-IID)data.To solve the problem,a graph clustering-based personalized federated learning method is proposed,which is called Graph clustering-based Personalized Federated Learning(GraphPFL).Firstly,each end device performs pre-training process locally.The current common similarity measurements are not effective in measuring the similarity of model parameters in tensor forms.Thus,the end devices will first convert the model parameters into weighted graphs according to the proposed method,and then upload the weighted graphs of model parameters to the edge server.Considering the relationship among the converted weighted graphs of model parameters from end devices with non-IID data,a personalizationenhanced features extraction method for end devices clustering is proposed.The experimental results show that GraphPFL not only significantly improves the average accuracy of FL,but also has better performance in terms of model personalization.(3)Research on a reliable federated learning method based on reputationIn a complex heterogeneous EC environment,end devices are easily corrupted by adversaries to launch poisoning attacks,and upload maliciously modified model parameters to reduce the accuracy of the FL and destroy the reliability of the model.To solve the problem,a reputation-based reliable federated learning method is proposed,which is called Reputation-Based Reliable Federated Learning(RepRFL).Firstly,the reputation values of the end devices that want to participate in FL are calculated according to the reputation model before the task starts,and the end devices are primed according to the reputation values.Secondly,an end device hiding mechanism is proposed to avoid the situation that end devices with high reputation values are more vulnerable to attacks.To circumvent the negative impact of dishonest end devices joining the model training on the global model,an aggregation strategy based on the elite campaign is proposed.Finally,to encourage more end devices to participate in FL honestly and strenuously,an incentive mechanism is designed.Security analysis and experimental results show that RepRFL not only achieves incentives for end devices,but also ensures the reliability of FL in the case of malicious end devices sending poisoning attacks.(4)Research on a fair federated learning method based on punishment mechanismIn FL oriented towards ring architecture,dishonest end devices may send wrong model parameters to harm the model performance of other end devices during model training,and selfish end devices may also leave maliciously after getting useful intermediate information,thus destroying the fairness of FL.To solve the problem,a fair federated learning method based on punishment mechanism is proposed,which is called Punishment-based Fair Federated Learning(PunFFL).Firstly,before the start of FL,all end devices need to pay a specified amount of upper deposit and lower deposit according to the rules.All transactions are recorded on the blockchain to ensure the security of transactions and digital currency.Then end devices perform model training,and the Hash algorithm is used to ensure the correctness of the transmitted models.The punishment mechanism ensures that all end devices either get the final result of FL or the honest end devices are compensated with digital currencies,thus achieving fairness.Security analysis and experimental results show that PunFFL not only guarantees the correctness of model parameters transmitted by end devices,but also ensures the fairness of FL in the case of malicious end devices launching escape attacks.