A CNN-transformer hybrid approach for an intrusion detection system in advanced metering infrastructure

Bi-directional communication networks are the foundation of advanced metering infrastructure (AMI), but they also expose smart grids to serious intrusion risks. While previous studies have proposed various intrusion detection systems (IDS) for AMI, most have not comprehensively considered the impact of different factors on intrusions. To ensure the security of the bi-directional communication network of AMI, this paper proposes an IDS based on deep learning theory. First, the invalid features are eliminated according to the feature screening strategy based on eXtreme Gradient Boosting (XGBoost), after which the data distribution is balanced by the adaptive synthetic (ADASYN) sampling technique. Next, multi-space feature subsets based on the convolutional neural network (CNN) are constructed to enrich the spatial distribution of samples. Finally, the Transformer is used to construct feature associations and extract crucial traits, such as the temporal and fine-grained characteristics of features, to complete the identification of intrusion behaviors. The proposed IDS is tested on the KDDCup99, NSL-KDD, and CICIDS-2017 datasets, and the results show that it has high performance with accuracy of 97.85%, 91.04%, and 91.06% respectively.