Internet of Things Intrusion Detection System based on Transfer Learning

Abstract

Listen

Translate article

Because of the many types of Internet of Things (IoT) attacks, as well as the lack of computing resources for some devices, the small number of anomalous datasets and the lack of updated data, resulting in insufficient training data and computational power for supervised learning-based intrusion detection models and the Internet datasets still dominate the research on IoT intrusion detection. Therefore, this paper proposes a VGG-RED approach to IoT intrusion detection based on Transfer Learning (TL), which migrates the feature weights from the training of two Internet intrusion detection datasets to the training of three IoT intrusion detection datasets respectively, and inter-migrates the three IoT datasets. The classification models are trained by model parameter migration and neural network fine-tuning. Unlike previous work on extracting manually designed features, this method retains the end-to-end learning performance of Deep Learning (DL), reduces the risk of concept migration occurring, and reduces human intervention. Experimental results demonstrate the feasibility of the VGG-RED model to migrate Internet attack classification to the IoT. The model can effectively improve the accuracy of malicious attack detection in IoT environments while reducing computational resources, and has a strong generalization capability, with an accuracy improvement of about 2% and a time reduction of 7-13%.