Federated Learning Framework For Balancing Job Completion Time And Model Accuracy

Mobile devices and Internet-of-Things(Io T)technology have led to increased data dispersion across devices and systems,containing valuable information that is not directly usable for ML models training due to data privacy and security issues.To address these issues,Federated Learning(FL)has been proposed.FL is a machine learning(ML)paradigm that protects data privacy while using multiple devices to train a global model.FL has gained research interest and is widely used in various fields,with varying requirements.FL’s expanding application areas have brought to light the varying requirements for FL tasks in different fields.For example,FL tasks such as speech recognition on Gboard require quick feedback,while medical diagnosis systems require precise judgments.These tasks have varying preferences for completion time and model accuracy,making it a worthy research question to develop different training strategies that maximize FL training accuracy.This paper proposes a Federated Learning framework called Hca,which aims to balance the time and accuracy requirements and develop efficient training algorithms based on their preferences.The framework divides the FL training process into three steps.Firstly,deduce the optimal parameter set for training through the method of estimation,and flexibly adjust the training strategy according to different task requirements.Secondly,a novel device selection algorithm is designed to select the devices with the best historical contributions and data effects from clients for participation,thereby improving the training efficiency.Finally,the reduction of model loss over consecutive rounds is used as a weighted factor for integrated calculation,further enhancing the training efficiency.The framework’s effectiveness is validated through theoretical analysis and extensive testbed experiments adopting FAVOR federated learning platform.The results show that compared to two state-of-the-art FL frameworks,Hca can increase task completion time by up to 34%,improve model accuracy by up to 9.1%,and reduce communication rounds for FL by up to 75%.This paper proposes an intelligent algorithm,Hca RL,based on Deep Reinforcement Learning(DRL)to tackle the challenge of quantifying and interrelated multiple performance factors.The proposed algorithm can improve training performance by selecting the best client devices while considering multiple performance factors.Extensive testbed experimental results demonstrate that Hca RL further improves performance based on Hca,including a 5% increase in model accuracy and a 23.6% reduction in completion time over Hca.