A Dual Channel Attention Mechanism-Based Intrusion Detection Model for Advanced Metering Infrastructure

Advanced metering infrastructure (AMI) has been widely used in the smart grid, but the interactivity and edginess of AMI make communication networks subject to cyber-attacks or intrusion threats, which bring severe security risks to the smart grid. The existing researches on handling the imbalance of AMI communication traffic data are inadequate, ignoring the complex nonlinear interaction relationship, unable to fully utilize time and spatial information, and are unsuitable for learning multidimensional features, resulting in low detection and high false alarm rates. Addressing the aforementioned issues, we propose an intrusion detection scheme rooted in deep learning to identify various attacks. Firstly, we utilize the embedded methods for feature selection and employ adaptive synthetic sampling (ADASYN) to augment the sample size of certain classes. Secondly, we design a dual-channel network for precise intrusion detection. This parallel network integrates convolutional neural networks (CNN) and bidirectional gate recurrent units (BiGRU) to capture spatial and temporal features. Furthermore, the attention mechanism is utilized for distributing feature weights across both channels, enhancing convergence speed and classification model accuracy. Finally, experimental findings using the KDDCup99 and NSL-KDD datasets demonstrate that the proposed model achieves high performance, attaining accuracies of 99.97% and 99.91%, respectively.