Mitigating Generator Bias in Low-Participation Model-Heterogeneous Federated Learning

Model heterogeneity and limited client participation are two defining challenges in real-world federated learning (FL). Representative generator-based distillation methods, such as FedGen and DFRD, use a server-side generator to synthesize pseudo-data for knowledge aggregation. However, under sparse participation, the generator suffers from biased supervision, overfitting to frequent participants and overlooking rarely seen clients, which we refer to as generator bias. To address this, we propose FLEdge (Federated LEarning with Debiased GEnerator), a unified framework that mitigates generator bias and supports robust personalization under heterogeneous models. FLEdge introduces Debiased Knowledge Distillation (DKD), which reweights both client- and sample-level contributions during distillation to correct distributional imbalance and guide the generator toward more comprehensive pseudo-data. To further enhance downstream personalization, we propose Bidirectional Contrastive Personalization (BCP), which decomposes local models into encoders and classifiers, enabling contrastive alignment with global features and temporal consistency with past decision boundaries. Extensive experiments on popular benchmarks show that FLEdge consistently improves low-participation model-heterogeneous FL, achieving up to 3.86% absolute accuracy gains over the best competing baseline on Tiny-ImageNet. Diagnostic analysis further shows that DKD reduces the empirical distillation loss gap from 1.239 to 1.084, providing empirical diagnostic evidence of reduced generator bias.