Haroon R Lone

FedPrism dynamically adapts to non-IID data in federated learning by decomposing client models into global, group, and private components, outperforming traditional aggregation methods.

Federated learning models can be significantly compressed and made more fair across clients by pruning different parts of the model for different clusters of users, without sacrificing accuracy.

Personalized federated learning can unlock accurate cross-country mood inference from smartphone data, even with heterogeneous sensing modalities and privacy constraints.