When Generator Replay Degrades: Projected Rehearsal Orchestration for Heterogeneous Federated Class-Incremental Learning

The article presents PRO, a novel framework for Federated Class-Incremental Learning (FCIL) that replaces synthetic input replay with projected rehearsal orchestration, addressing challenges posed by heterogeneous task streams and uneven supervision. The framework maintains compact class-level projected memories on the server and enables clients to conduct balanced multi-task training, while the enhanced PRO-MAX variant incorporates neighborhood-weighted memory alignment to mitigate representation drift. Experimental results demonstrate that both PRO and PRO-MAX significantly improve knowledge retention and performance across diverse benchmarks, highlighting the importance of replay quality over quantity in FCIL scenarios.