<ul data-eligibleForWebStory="true"><li>Federated learning (FL) enables multiple data holders to train a ML model without sharing data externally.</li><li>FL involves workers updating a model locally and sharing their gradients with a central server.</li><li>Byzantine-resilient FL prevents malicious participants from harming model convergence.</li><li>Common strategies in FL ignore outlier gradients to thwart attacks.</li><li>In heterogeneous data settings, distinguishing outliers is challenging.</li><li>A new approach, Worker Label Alignement Loss (WoLA), aligns honest worker gradients in heterogeneous data settings.</li><li>WoLA helps in identifying malicious gradients and outperforms existing methods in such settings.</li><li>The paper includes theoretical insights and empirical evidence supporting WoLA's effectiveness.</li></ul>

Weighted Loss Methods for Robust Federated Learning under Data Heterogeneity

Discover more