BESplit Mitigates Non-IID Bias in Split Federated Learning

The BESplit framework uses Evidential Aggregation to reweight client contributions based on uncertainty in Split Federated Learning. This architecture-aware approach compensates for biased optimization and unstable convergence caused by non-IID data distributions.