Reliability-Aware Generator Consolidation with Feature-Gated Adversarial Replay for Federated Class-Incremental Learning

Federated class-incremental learning must retain earlier knowledge while combining clients with different class distributions and task histories. We study synchronized stage transitions with client-specific task orders and spatial label skew. FedRUA organizes replay into two server stages. First, a conditional generator learns from cached client classifiers whose real-data histories include the requested class. Predictive disagreement attenuates multi-expert supervision, while a singleton fallback combines predictive confidence, feature-statistic consistency, and update deviation. Second, the server fine-tunes the global classifier using weighted clean replay and feature-gated adversarial replay. The gate controls the additional adversarial loss while retaining clean replay. Clients use the frozen shared generator for local replay without retaining historical raw examples. On five-task CIFAR-100, FedRUA improves mean final accuracy by 0.88 percentage points over DGR-FCL and reduces mean forgetting by 1.70 points. Under Dirichlet label skew with $\alpha =0.5$, the accuracy margin increases to 1.45 points. On $64\times 64$ ImageNet-1K, the corresponding accuracy and forgetting margins are 0.30 and 0.70 points. These results support the integrated procedure for a trusted server under the evaluated task-order and spatial heterogeneity.