| With the advent of Internet of Things(Io T),an exponential growth of data are generated at edge devices.However,various laws and regulations such as Personal Information Protection Law may be a hindrance to direct data sharing.Moreover,the competitive relationship between different organizations also prevents data from being shared.This phenomenon is called a data island.To overcome the problem of isolated data islands,federated learning(FL)was proposed by Google in 2016.In FL,edge devices(i.e.,clients in FL)train the model with local data,and cloud server learns shared model by aggregating locally computed updates with privacy guarantee.However,heterogeneous properties of devices may cause a long single round duration in synchronous federated learning and stale gradient of clients in asynchronous federated learning.To solve these issues,we propose a buffered semi-asynchronous federated learning with efficient client selection strategy,named BSACS-FL.First of all,considering the bottleneck of communication resources,we only select K clients to upload their trained models in each round.To select the proper clients,We split the staleness of the client’s local model into Training Staleness(TS)and Waiting Staleness(WS).It has been proved that the staler the local model the more damage to the global model,so we select the latest model first.To accelerate the training process,if the freshness of the model is the same,we choose the clients with stronger computing capability.In addition,the computing resources are always changed because of the different applications running on the devices.So we use Upper Confidence Bound(UCB)-based MAB mechanism to dynamically predict which round the clients will complete the training.In order to further weaken the damage to the global model caused by device system heterogeneity and data heterogeneity,the server reweights the local model according to its staleness and similarity to the global model.At last,we set a buffer on client side,which is used to store models that have been trained by clients but not been selected by the server to aggregate.We conduct extensive experiments with realistic human activity dataset Mobi Act and image recognition dataset MNIST.First,the optimal combination of hyperparameters for global model aggregation is determined through experiments,and then the impact of different sizes of aggregation time slices on the algorithm is analyzed.At last,compared with synchronous federated learning algorithm Fed Avg,asynchronous federated learning Fed Async,and semi-asynchronous federated learning algorithm Fed Buff,the results corroborate that our algorithm can achieve superior performance.As far as we know,in the same kind of research,BSACS-FL is the first algorithm that sets the buffer structure on the client side,and is the first to predict training staleness of the local model instead of directly predicting the exact training ability of the client,which improves the fault tolerance of the prediction results.This paper finally verified the effectiveness of BSACS-FL through experiments. |
