FedDS: Data Selection for Streaming Federated Learning with Limited Storage

Federated learning (FL) is a privacy-preserving distributed learning framework where model training is performed locally on distributed devices. Unlike traditional FL, which assumes a fixed local dataset, this paper focuses on the more realistic scenario of FL with streaming data. In Streaming Federated Learning (SFL), new data continuously arrives over time, and due to the limited local storage capacity on devices, some data is inevitably discarded. The discarded data may be forgotten by the model, leading to a decline in model accuracy. To this end, we introduce Federated Data Slimming (FedDS), a data selection scheme designed to determine which data should be stored locally. Particularly, FedDS considers both gradient norms and directions when making data selections. We evaluate the performance of FedDS against several previously proposed schemes using various datasets. Our experimental results demonstrate that FedDS surpasses all baseline schemes, achieving the fastest convergence rate and the highest test accuracy.