TGDCLNet: Teacher-Guided Denoising Contrastive Learning Network-Based IoT Network Intrusion Detection

With the rapid development of Internet of Things (IoT) technology, many devices are connecting to networks, making the security risks of IoT devices a growing focus of concern. The demand for enhancing the accuracy of IoT intrusion detection is becoming increasingly urgent. In the face of uneven traffic distribution and the problem of unknown attack types in the network flow data collected by IoT devices, traditional supervised learning models are not robust enough in practical applications. To address these issues, we propose an innovative teacher-guided denoising contrastive learning network (TGDCLNet), which employs a dual autoencoder structure for teacher and student models to perform network intrusion detection. In preprocessing, we innovatively design hybrid rebalancing strategies to adapt to different data distribution scenarios and select the best strategy combination. In the denoising learning module, we propose a denoising multidimensional balanced mean squared error loss function to reduce the impact of noise in the collected traffic on the training of both teacher and student models. Furthermore, we design a teacher guidance module, which takes label information as a feature input into the teacher autoencoder, and propose an equilibrium-weighted binary cross-entropy loss function to guide the learning process of the student model. The experimental results on six representative datasets show that our method outperforms other supervised contrastive learning algorithms. We also verify the robustness of our method in the complex and dynamic real-world IoT environment. Our method performs better in challenging environments with unknown attack types and scarce traffic data than other contrastive learning methods, providing a stable and effective solution for IoT intrusion detection.