Abstract
With the rapid advancement of Internet of Things (IoT) technology, intrusion detection systems (IDSs) have become pivotal in ensuring network security. However, the data produced by IoT devices is typically sensitive and tends to display non-independent and identically distributed (Non-IID) properties. These factors impose significant limitations on the application of traditional centralized learning. In response to these issues, this study introduces a novel IDS framework grounded in federated learning and knowledge distillation (KD), termed FD-IDS. The proposed FD-IDS aims to tackle issues related to safeguarding data privacy and distributed heterogeneity. FD-IDS employs mutual information for feature selection to enhance training efficiency. For Non-IID data scenarios, the system combines a proximal term with KD. The proximal term restricts the deviation between local and global models, while KD utilizes the global model to steer the training process of local models. Together, these mechanisms effectively alleviate the problem of model drift. Experiments conducted on both the Edge-IIoT and N-BaIoT datasets demonstrate that FD-IDS achieves promising detection performance across multiple evaluation metrics.