Proceedings Article | 22 February 2023
KEYWORDS: Education and training, Machine learning, Sampling rates, Data privacy, Performance modeling, Stochastic processes, Data processing, Data modeling, Computer simulations, Mathematical optimization
Federated Learning (FL) protects users’ privacy by only uploading the training result instead of gathering all the private data. However, achieving the desired model performance often requires a large number of iterations of parameter transfer between client and central server. Currently, selecting a fixed number of clients to participate in training can slightly reduce the communication overhead during model training, but ignore the impact on model training accuracy. In this paper, we propose an adaptive chosen client number K scheme, which can give a better tradeoff between accuracy and cost. Firstly, through experiments, we find that increasing extracted clients’ number K can reduce iterations’ number T, but after K increases to a certain extent (𝐾1), T will no longer reduce significantly. Similarly, increasing K can further improve the accuracy of model training, but K is large enough (𝐾2 ≥ 𝐾1), the accuracy will also no more be improved remarkably. Thus, [𝐾1,𝐾2] is the optimal range. Secondly, we conduct experiments on different datasets with different number of clients, and find that the optimal client’s number growth rate q’ for different conditions is 0.02. According to the experimental results, we set the initial K to be 𝐾1 for the optimal T, when the model update magnitude in two adjacent iterations is less than a threshold, the number of clients participating in training will increase by q’ to speed up the convergence until K reaches K2, otherwise it will remain unchanged. Finally, we use our algorithm to improve present FL algorithms. Through experiments, we demonstrate that our algorithm is suitable for existing differential private FL algorithms.
