Research And Application Of A Constructive Federated Learning Method

The rapid development of mobile Internet technology has enriched everyone’s spir-itual and cultural life to a large extent.As its core means of production,data has gradually attracted the attention of all walks of life.Social problems caused by privacy leaks and data trading are not uncommon.How to implement artificial intelligence algorithms on the premise of protecting citizens’privacy and break the dilemma of data islands is the direction of the joint efforts of academia and industry today.Federated learning is one of the effective ways to solve such problems.In the framework of Federated learning,this paper proposes a communication effi-cient distributed computing method.By constructing KKT condition of?₀constrained minimization problem,the parameter estimation and feature selection of linear model in high dimensional sparse case are realized by gradient descent method.The inno-vation of this paper is reflected in two aspects.First,the algorithm makes full use of the gradient information of the data,so that the original data can complete the joint modeling without going out locally,which ensures the security and privacy of the data;Second,the algorithm is based on high-dimensional sparse settings,and uses the method of first determining the support set and then estimating the parameters to reduce the complexity of the algorithm and improve the computational efficiency.Experimental results show that this algorithm can effectively carry out distributed computing,and compared with other commonly used distributed methods,the convergence speed is faster,the communication cost is lower,and the calculation results are more accurate.In addition,compared with the classical penalty method,this algorithm has the char-acteristics of high precision and high credibility in feature selection for medium sparse problem in single machine system.