TransIDS: A Transformer-based approach for intrusion detection in Internet of Things using Label Smoothing

Abstract

As Internet of Things (IoT) is rapidly developing and popularizing, IoT devices generate a large amount of network traffic information, which requires reliable IoT traffic intrusion detection techniques to continuously improve network security mechanisms. Most existing intrusion detection methods based on machine learning and deep learning in IoT rely on complex feature reduction and feature selection techniques, and the models cannot focus on important features adaptively and have poor global modeling capability for high-dimensional sequential features. To overcome the dependence on feature preprocessing, realize the model adaptively focus on important features, and further enhance the ability to extract global features, this paper proposes a Transformer-based IoT intrusion detection method called TransIDS. TransIDS has remarkable global modeling capability by introducing a multi-headed self-attention mechanism, which can extract multiple global temporal features, adaptively adjusting the attention to high-dimensional features. To overcome the undesirable effects of unbalanced datasets, we adopt Label Smoothing to add noise to sample labels to avoid over-reliance on training samples, which can enhance the generalization ability of the model. Finally, the performance of the proposed method is verified on a TON-IoT standard dataset in a real environment, and the experimental results show that the proposed method achieves the superior recognition performance compared to other advanced methods. In addition, we also investigate the effect of hyperparameters on the detection performance.