FEELPGen: Data-Free Knowledge Distillation for Personalized Federated Learning Across Heterogeneous Edge Silos

Deploying machine learning models on large-scale IoT devices in edge networks is challenging. Federated edge learning (FEEL) has emerged as a potential solution based on a hierarchical architecture. However, existing research relies primarily on an idealized cross-device assumption, overlooking more realistic cross-silo scenarios where devices typically belong to different organizational silos. To facilitate multi-group collaboration, we first propose a semi-decentralized FEEL structure called FEELPGen, in which different silos collaborate in training to maximize local model benefits without relying on trusted third-party coordination. Based on that, a two-layer aggregation algorithm is proposed to enhance the generalization ability under highly heterogeneous data distribution. For inner-silo learning, we devise a heterogeneity-aware, synchronous inner-silo aggregation algorithm utilizing data-free knowledge distillation based on generative learning (Gen). Feature vectors are generated to approximate silo knowledge. For inter-silo learning, a personalized (P), asynchronous inter-silo aggregation algorithm is proposed with adaptive selection and dynamic weight queues. To further improve efficiency, we introduce an optional optimized scheme, FEELPGen+, which integrates a privacy-preserving dimension-reduction algorithm. Finally, we provide a detailed analysis for convergence and complexity to verify the feasibility of FEELPGen. Extensive experiments demonstrate that FEELPGen achieves significant improvement in accuracy compared to the state-of-the-art schemes.