In the Internet of Things (IoT), a large number of devices generate data. Federated learning can take advantage of the distributional nature of these devices to parallelize model training, thus speeding up model training. The data used for model training in IoT can be heterogeneous, which can greatly reduce the inference accuracy of the global model. To improve the accuracy of models, we propose a new scheme: for the knowledge transfer process, we utilize soft lables to transfer knowledge between clients, and for the parameter aggregation process, we group all the clients based on the similarity of their soft lables and select one client from each group for aggregation in each round of aggregation, and design its weight when participating in the aggregation based on the reliability of its soft lables. We demonstrate through experimentation that our approach is effective and outperforms previous algorithms, resulting in improved model accuracy.

在物联网 (IoT) 中，大量设备生成数据。联邦学习可以利用这些设备的分布式特性来并行化模型训练，从而加快模型训练速度。物联网中用于模型训练的数据可能是异构的，这会大大降低全局模型的推理精度。为了提高模型的准确性，我们提出了一种新方案：对于知识传输过程，我们利用软标签在客户端之间传输知识，对于参数聚合过程，我们根据软标签和参数的相似性对所有客户端进行分组。每轮聚合从每组中选择一个客户端进行聚合，并根据其软标签的可靠性设计其参与聚合时的权重。我们通过实验证明我们的方法是有效的并且优于以前的算法，从而提高了模型的准确性。