<ul><li>DRDM is a novel algorithm introduced for distributed learning to address challenges in heterogeneous environments by combining distributionally robust optimization with dynamic regularization to minimize client drift.</li><li>DRDM optimizes a min-max objective function to maximize performance for the worst-case client, aiming to achieve fairness and robustness in model performance.</li><li>The algorithm leverages dynamic regularization and efficient local updates, reducing the number of communication rounds required for training.</li><li>Extensive experiments on benchmark datasets show that DRDM improves worst-case test accuracy and requires fewer communication rounds compared to existing state-of-the-art approaches.</li></ul>

Distributionally Robust Federated Learning with Client Drift Minimization

Discover more