CRAFT: Conflict-Resolved Aggregation for Federated Training

Researchers introduced CRAFT, a federated learning aggregation method designed to resolve conflicts among client updates.

CRAFT reframes aggregation as a geometric correction problem, deriving a closed-form solution with layer-wise adaptation.

The method improves global model accuracy while reducing performance disparities across clients on heterogeneous data.

It offers a more stable alternative to naive averaging for non-IID federated training.