Design And Implementation Of Multi-level Federated Learning Optimization Algorithm In Edge-End Collaborative Environment

With the rapid development of information technology,data is stored in various industries in a fragmented form,and data islands are formed due to business competition and industry barriers.At the same time,data privacy and security are becoming more and more important.How to break the data silands under the premise of protecting data privacy,and efficiently collaborate these data has become a key problem to be solved urgently.Federated learning technology is a distributed machine learning technology that realizes data privacy protection.By uploading local model information instead of original data,understanding the value of aggregated data from local knowledge,it provides a targeted solution to the above problems.However,in the complex edge-end collaborative environment,massive devices bring a huge burden to the federated learning center server,and even if the federated learning does not transmit the original data during the training process,the model parameters or gradients still have the risk of leaking user privacy.In addition,due to different user behavior habits and other reasons,the data saved on each device realitily is often not independent and identically distributed,which will cause the problem of slow model convergence and performance degradation during the training process.Aiming at the above problems,this thesis designs and implements a multi-level federated learning optimization algorithm in a edge-end collaborative environment.It mainly includes the following two parts:(1)In order to reduce the burden of the central server in the edge-end collaborative environment and further protect user privacy,this thesis proposes a multi-level federated learning mechanism integrating differential privacy.First,a multi-level federated learning mechanism is designed to decouple the central server from large numbers of local devices,effectively reducing the burden on the central server.Furthermore,a locally adaptive differential privacy algorithm adapted to the multi-level federated learning mechanism is proposed.By adding a dynamically adjusted Laplace distributed random number to the model parameters for encryption,malicious attackers cannot obtain the original data reflecting the data distribution.At the same time,different degrees of encryption are performed according to the performance of the nodes to ensure the model availability of excellent nodes,thereby improving the model effect.Finally,experiments show that the proposed mechanism can achieve more efficient and reliable federated learning tasks in the edge-end collaborative environment.(2)Aiming at the problems of slow model convergence and performance degradation caused by non-ⅡD distribution of data on local devices under the multi-level federated learning mechanism,this thesis proposes a node selection algorithm based on reinforcement learning.According to the architectural characteristics of the multi-level federated learning mechanism(cloud center-edge server-client)for analysis and design,and select better nodes through the learning and analysis of intermediate model parameters to offset the impact of non-ⅡD data.Finally,it is proved by experiments that the node selection algorithm based on reinforcement learning can speed up the convergence of the model and improve the performance of the model on the experimental data set.The experimental results show that the multi-level federated learning mechanism integrating differential privacy and the node selection model based on reinforcement learning proposed in this thesis can effectively solve the above problems,realize efficient and reliable data collaborative sharing in the edge-end collaborative environment,and further promote the practical application of federated learning technology.