A deep residual SConv1D-attention intrusion detection model for industrial Internet of Things

The development of intrusion detection technology has contributed greatly to industrial Internet of Things (IIoT) security. However, intrusion detection system (IDS) for IIOT anomaly detection suffer from limited computational costs. In addition, the problem of class imbalance in IIOT network traffic is a challenge for IDS. To this end, this paper proposes a deep residual SConv1D-Attention model, which improves the accuracy of detecting minority classes, increases the speed of model anomaly detection, and reduces the computational cost of the model. Specifically, we use a binary Particle Swarm Optimization (bPSO) algorithm to select the features of the samples and remove the redundant features of the samples, which improves the performance of the model. We design a novel SConv1D-Attention module that employs a one-dimensional version of depth separable convolution and self-attention for information integration, the computational cost is reduced while information loss is effectively minimized. In response to data imbalances, during training, we design a robust model loss function to increase the weight of the minority class and balance attention to learning in a few categories. We used the ACC, DR, FPR, Precision and F1_score indicators based on CICDDoS2019, NSL-KDD and X-IIoTID datasets to evaluate our model. The experimental results show that the binary and multiclassification results of our model reached 99.86%–99.99% and 99.42%–99.91%, respectively, on the basis of the ACC, DR, Precision and F1_score indicators of each dataset, and 0.03%–0.16% and 0.02%–0.42%, respectively, on the FPR indicators, which are superior to those of traditional deep learning methods and state-of-the-art models. High evaluation results show that our model can improve the efficiency of network intrusion detection. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.