Federated learning (FL) enables collaborative model training without sharing raw data. However, in real-world applications, clients often exhibit statistical heterogeneity, missing classes, and long-tailed distributions, which can substantially degrade the generalization performance of conventional parameter aggregation and some personalization approaches. Moreover, distillation or alignment-based methods may suffer from unstable supervision and difficult optimization under highly heterogeneous settings. To this end, this paper proposes a novel method called FKD-RTM (Heterogeneous Federated Knowledge Distillation Based on Residual-Enhanced Tree-to-MLP Knowledge Transfer). The key idea is to decouple local teaching from globally aggregatable student learning: we introduce a Gradient Boosting Decision Tree (GBDT) as a local teacher at each client, providing more reliable soft supervision based on shared feature representations for a Multi-Layer Perceptron (MLP) student that supports efficient global aggregation and adaptation. To further correct the prediction bias left after first-stage distillation, we introduce a residual enhancement mechanism. It learns complementary knowledge in the pre-normalization score domain and enables second-stage corrective learning. In addition, FKD-RTM performs partial parameter fine-tuning of the feature extractor and student model for personalized local adaptation. Personalized updates are excluded from global aggregation to avoid contaminating the global model. Experiments on multiple datasets, including CIFAR-100, demonstrate that the proposed FKD-RTM method consistently improves accuracy and generalization under diverse complex data settings and achieves a better trade-off between global and personalized performance.
Sheyun et al. (Thu,) studied this question.