【摘要】 联邦学习算法通常面临着客户端之间差异巨大的问题，这些异质性会降低全局模型性能，文章使用知识蒸馏方法缓解这个问题。为了进一步解放公共数据，完善模型性能，文章所提的DFP-KD算法使用无数据方法合成训练数据，利用无数据知识蒸馏方法训练鲁棒的联邦学习全局模型；使用ReACGAN作为生成器部分，并且采用分步EMA快速更新策略，在避免全局模型灾难性遗忘的同时加快模型的更新速率。对比实验、消融实验和参数取值影响实验表明，DFP-KD算法比经典的无数据知识蒸馏算法在准确率、稳定性、更新速度方面都更具优势。更多还原

【Abstract】 Federated learning algorithms usually face the problem of huge differences between clients,and these heterogeneities degrade the global model performance,which are mitigated by knowledge distillation approaches.In order to further liberate public data and improve the model performance,DFP-KD trained a robust federated learning global model using datad-free knowledge distillation methods;used ReACGAN as the generator part;and adopted a step-by-step EMA fast updating strategy,which speeded up the update rate of the global model while avoiding catastrophic forgetting.Comparison experiments,ablation experiments,and parameter value influence experiments show that DFP-KD is more advantageous than the classical data-free knowledge distillation algorithms in terms of accuracy,stability,and update rate.更多还原