在线异步联邦学习的客户优化选择与激励.

Federated learning enables different clients to collaborate and train a shared model while preserving user privacy. Motivating high-quality clients to participate in federated learning is crucial. In online federated learning environments, where clients join and leave training dynamically, evaluating and selecting clients in real-time poses a challenge. To address this challenge, this paper proposed an online federated learning incentive algorithm to optimize client selection and budget allocation, thereby enhancing the performance of federated learning under budget constraints.The proposed algorithm divided the budget into stages and computed optimal quality density thresholds based on historical sample information. The main idea was to dynamically assess the quality of client models and employ a quality threshold admission mechanism while limiting the number of participating clients. In theory, this paper proved that the incentive algorithm satisfied incentive compatibility, budget feasibility, and individual rationality. Experimental results demonstrate that the proposed online incentive algorithm achieves good performance in scenarios with different proportions of free-riding clients. Specifically, compared to existing methods, it achieves approximately 4% and 10%improvements on the EMNIST-B and CIFAR-10 datasets, respectively, under sufficient budget and in the presence of free-riding and mislabeled clients. [ABSTRACT FROM AUTHOR]