FREQSEL: A Frequency-Based Method for Client Selection in Federated Learning

Federated learning (FL) is a machine learning approach that allows multiple devices to collaborate in constructing a shared prediction model while keeping their sensitive data decentralized, avoiding the need for transmission to a central server. In each training iteration, only a subset of clients is involved. While this approach offers significant advantages, one of the primary challenges arises from the Non-Independent and Identically Distributed(N-IID) nature of the data across the participating clients. This can lead to bias when aggregating the parameters of diverse local models, ultimately reducing the global model’s accuracy. In this paper, we propose FREQSEL, an innovative federated learning algorithm for client selection designed to mitigate the accuracy drop caused by N-IID data across client devices. In FREQSEL, the server begins the communication process by prompting the clients, who then report the size of their data samples. Subsequently, the server calculates the class frequencies for each client, as well as those for the entire population, without accessing the clients’ datasets. These frequency distributions serve as the basis for client selection. Finally, the server distributes the global model to the chosen clients. By considering the imbalanced data distribution among clients, FREQSEL enhances model performance in heterogeneous data environments. Although sharing frequency vectors may raise privacy concerns, our method only exchanges aggregated statistical information, not raw or individual-level data, which significantly mitigates the risk of exposing sensitive information. We evaluate the effectiveness of our approach through simulations using publicly available datasets, and the results demonstrate that FREQSEL outperforms existing methods in terms of convergence speed and test accuracy.