<p>Federated Learning (FL) has become an important method for training powerful machine learning models within the Internet of Things (IoT) ecosystem. However, the data heterogeneity and scarcity (few-shot samples) in IoT present significant challenges, undermining FL’s effectiveness in achieving optimal model performance. Most existing methods address the heterogeneity challenge merely by constraining local model updates, neglecting the generalization limitations caused by simple aggregation of the global model. Therefore, we propose a FL method based on data-free knowledge distillation (DFKMD) and distribution shift, called FEDKDG, to enhance the effectiveness of the global model aggregation strategy and global knowledge transfer. Specifically, FEDKDG employs a lightweight conditional generative network trained on clients and uploaded to the server to produce high-quality synthetic data for DFKMD, effectively reducing communication overhead while addressing the server’s inability to access clients’ private data. In addition, a pre-trained large teacher model is integrated into the server aggregation phase to provide stronger and more comprehensive global knowledge. Furthermore, we introduce a distribution-shift-based data augmentation strategy with a contrastive learning design to mitigate the shift between global and local knowledge, allowing clients to better utilize the distilled global knowledge. Extensive experiments demonstrate that, compared to FL baselines, FEDKDG exhibits stronger competitiveness and superiority in addressing data heterogeneity challenges.</p>