BESplit: Bias-Compensated Split Federated Learning with Evidential Aggregation

In three linesBESplit introduces a Split Federated Learning framework to mitigate non-IID data effects. The method combines Evidential Aggregation for client contribution reweighting, Bias-Compensated Collaboration for representation alignment, and Dual-Teacher Distillation for model synchronization. Experiments on 5 benchmarks show improvements in accuracy and convergence stability.

Summary generated by Claude — human-verified