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

2026-06-14

AI Summary

This work addresses the challenges of knowledge forgetting and representation drift in heterogeneous federated class-incremental learning, which arise from inconsistent client task phases, label distribution shifts, and imbalanced supervision. To tackle these issues, the authors propose the PRO framework, which abandons conventional input-space replay and instead maintains a lightweight class-level projection memory at the server. A novel pseudo multi-task training mechanism enables clients to balance learning between current data and historical projection memory. The framework introduces, for the first time, a projection rehearsal orchestration strategy that operates without external pretraining. Furthermore, an enhanced variant, PRO-MAX, mitigates severe representation drift through neighborhood-weighted memory alignment. Extensive experiments demonstrate that the proposed approach significantly outperforms existing baselines across image, text, and graph data, maintaining stable performance and efficient knowledge retention even as the replay budget increases.

Abstract

Federated class-incremental learning (FCIL) becomes substantially harder when clients observe different label subsets, progress through tasks at different stages, and provide uneven supervision for the same semantic concepts. Existing FCIL methods often preserve old knowledge through input-space synthesis, but they can be fragile under heterogeneous task streams and difficult to transfer across modalities. To alleviate such issues, we propose PRO, a framework that replaces synthetic input replay with projected rehearsal orchestration. To remove external pretraining, we evaluate all methods under the same warmup. After this, PRO maintains compact class-level projected memories on the server and allows clients perform balanced pseudo multi-task training over current examples and old projected memories. To handle stronger representation drift, we further introduce PRO-MAX, which augments PRO with neighborhood-weighted memory alignment while preserving the same server-light principle that the server only aggregates model updates and memory statistics. Across image, text, and graph benchmarks, PRO and PRO-MAX improve retention and final utility under heterogeneous streams while remaining competitive in homogeneous FCIL. Even when baselines are given expanded replay budgets, they degrade under supervision imbalance and stage misalignment, indicating that replay quantity alone does not resolve replay-quality failures. Additional weak-task diagnostics further show that larger replay mismatch is associated with larger downstream degradation, while our method keeps projected memories better aligned with the evolving representation.

Problem

• Federated Class-Incremental Learning
• Heterogeneous Task Streams
• Knowledge Retention
• Supervision Imbalance
• Representation Drift

Innovation

• Projected Rehearsal Orchestration
• Federated Class-Incremental Learning
• Heterogeneous Task Streams
• Memory Alignment
• Server-Light Framework