Efficient federated knowledge distillation strategy for heterogeneous data in the Internet of Things

With the rapid evolution of Internet of Things (IoT) technology and the advent of the big data era, a vast network of interconnected devices and sensors has emerged. Traditional centralized model training methods are insufficient for meeting the data privacy and sharing requirements of IoT systems. Moreover, the inherent heterogeneity of data generated by these devices poses a significant security challenge. In response, this paper presents an innovative federated knowledge distillation strategy tailored for handling heterogeneous data in IoT environments. The proposed approach incorporates knowledge distillation into the federated learning framework, facilitating the transfer of knowledge from a teacher model to a more efficient student model. Communication efficiency is enhanced by transmitting predictive labels instead of the entire model parameters. To further narrow the gap between the teacher and student models and improve generalization, Kullback-Leibler and regular terms are integrated into the federated learning objective function. Additionally, we introduce a CNN-Transformer hybrid model on the client side, comprising a classifier and discriminator network. This architecture improves the quality of predicted labels, addressing training inefficiencies associated with highly heterogeneous data.