Heterogeneous Fairness Algorithm Based On Federated Learning

With the rapid development of Io T technology,increasingly large amounts of data.The emergence of large amounts of data has led to the rapid development of technologies such as big data analysis and machine learning algorithms.The above technologies need to process the data in a centralized way,which can not ensure the privacy of the data.This is contrary to people's requirements for privacy protection.Based on this,a new paradigm of federated learning came into being.Federated learning sinks data training to the edge instead of centralized processing.The training model is sent to the client device through the central server for local training,and the training results are sent back to the central server.After multiple iterations until convergence,a better global model is finally obtained.On the one hand,federated learning solves the problem of privacy protection and data island of users.On the other hand,because the paradigm does not need to transmit a lot of data through the network,it greatly reduces the network pressure.Due to the characteristics of federated learning training on distributed devices,there are statistical heterogeneity and system heterogeneity differences among the training devices.The difference between devices directly leads to the performance difference between devices,and the quality of the trained local models are therefore different.In the process of using traditional federated learning for local training,devices with poor performance may have the risk of frequent offline or unable to occupy a dominant position in global aggregation.That is to say,it has little influence on the final global optimization objective.Under this premise,the trained global model is bound to deviate from these poor performance devices,so the trained global model has the risk of not fitting all the devices.In other words,the trained global model is unfair to the devices with poor performance.However,at present,scholars at home and abroad mainly optimize the fairness of federated learning based on the statistical heterogeneity difference,and there is no literature report on the system heterogeneity difference.In view of the lack of research in this aspect,this paper studies the fairness of federated learning under the condition of system heterogeneity.In response to the above problems,this paper proposes a federated learning fairness optimization goal OPT for system heterogeneity differences,Using three different parameters,the number of local samples of the devices,the number of local training rounds per unit time of the device,and the accuracy of local training per unit time of the device,design and obtain different device influence factors for each device.The device influence factor is introduced into the optimization target OPT to dynamically adjust the proportion of different performance devices in the optimization target.At the same time,this paper designs a system heterogeneous fair federated learning algorithm(SHFF).By dynamically adjusting the global fairness parameters,the algorithm can control the fairness optimization of different data sets according to actual needs.Finally,this paper designs a verification experiment.The experimental results show that the proposed SHFF algorithm on Vehicle,Synthetic and fmnist data sets,under the premise that the global average accuracy and the average accuracy of better performance devices have little fluctuation,the average accuracy of poor performance devices are significantly improved,and the variance is significantly reduced.Compared with the baseline algorithm q-Fed Avg,the average accuracy of SHFF algorithm proposed in this paper is improved by 25.2% on Vehicle Worst10%devices,and the variance decreased by 61.3%;the average accuracy is improved by46.5% on Synthetic Worst40% devices,and the variance decreased by 21.8%;the average accuracy is improved by 2.7% on fmnist Worst10% devices,and the variance decreased by 15.8%.Experiments prove that the SHFF algorithm is effective and universal.