Abstract
In the field of Industrial Internet of Things (IIoT), existing intrusion detection models face challenges in three main areas: low accuracy in detecting attack traffic, feature redundancy when dealing with high-dimensional and complex attack traffic, making it difficult to capture critical information, and a tendency to favor learning common categories while neglecting rare categories when handling imbalanced data. To tackle these challenges, this study introduces an intrusion detection method that combines an attention mechanism, Bidirectional Gated Recurrent Units (BiGRU), and Inception Convolutional Neural Network (Inception-CNN) to enhance the model's detection rate. Simultaneously, the method employs a mixed sampling strategy for data resampling to address the bias learning issue caused by data imbalance. Additionally, the method employs a hybrid sampling strategy for data resampling to address the bias learning issue caused by data imbalance. It also incorporates denoising techniques to handle potential dataset noise introduced by hybrid sampling. Furthermore, a feature selection method combining Pearson correlation coefficient and Random Forest is applied to eliminate feature redundancy, enhancing the model's ability to capture crucial information from high-dimensional attack traffic. Experimental validation on internationally recognized datasets (Edge-IIoTset, CIC-IDS2017, and CIC IoT 2023) affirms the reliability of the proposed intrusion detection method. This approach underscores the significance of intrusion detection in the security of Industrial IoT and showcases its potential in addressing pertinent challenges in network security.