Research On Federated Learning Based On Blockchain

With the advent of the era of big data and the growing demand for artificial intelligence applications,the advantages exhibited by machine learning make it an important tool for the development of intelligence in different fields.However,due to its centralized training model,”data silos” and data privacy issues become significant challenges for machine learning.As an emerging artificial intelligence technology,federated learning is distributed locally through terminals.It avoids the direct upload of private data,and ensures information security and privacy during big data exchange.Nevertheless,traditional FL still faces the problem of being tampered with due to long communication distance.The change of some parameters will cause the model accuracy to change or not converge,resulting in training errors or even system collapse.This thesis decentralizes the traditional FL based on the blockchain network,and collects the model parameters of each terminal through the blockchain network.As a safe and reliable distributed data transaction system,blockchain's trust mechanism and traceability ensure data correctness and privacy.It can effectively solve the tamper-proof problem of federated learning system.However,considering the differences in distributed terminals and the complex block-forming method of the blockchain,a single round of federated learning may generate huge latency,which affects the learning efficiency of FL.Therefore,how to optimize the federated learning delay based on blockchain is the key research content of this thesis.This thesis considers that in the case of large differences in terminal computing resources and communication environments,the overall integration of model information uploaded by all terminals will lead to huge latency.However,when the difference is small,the model information uploaded by each terminal is separately authenticated into blocks,which will generate additional network transmission delay and mining delay.Therefore,this thesis designs a mechanism for combining terminals into blocks,and makes a compromise between the two combining methods.This thesis mainly studies the optimal blockforming strategy of blockchain considering terminal resources,data distribution and network resources in the process of single-round federated learning,so as to minimize the delay of single-round federated learning.And a mixed integer non-convex optimization model is established.Through the comparison of numerical results,the block model and algorithm designed in this thesis can provide the optimal block strategy and reduce the training delay.Considering that blockchain nodes are edge nodes with certain computing resources,the thesis introduces a model splitting mechanism.Reduce variance between terminals and terminal computing energy consumption by computing offloading.Since the calculation offload will bring additional communication energy consumption and delay to the terminal,and the terminal energy is limited.In order to reduce the energy consumption of the terminal under the condition of guaranteeing the delay,this thesis considers the delay and computing resource constraints,designs a resource allocation strategy to minimize the system energy consumption,and establishes a mixed integer non-convex optimization model.In this thesis,a heuristic algorithm is designed to solve the model.By splitting the model into three sub-problems: computing resource allocation,terminal model splitting,and blockchain block node selection,the final solution is obtained after multiple iterations.The simulation results show that the resource allocation strategy given in this thesis can effectively obtain the resource allocation strategy of the network,so as to reduce the energy consumption of the terminal and improve the resource utilization rate under the condition of guaranteeing the delay.