Robust Intrusion Detection for Industrial IoT

Huaishuo Ren, Liangyi Gong, Xiuliang Mo, Lanqi Yang, Zuwei Yin

Abstract: Due to the significant role of Industrial Internet of Things (IIoT) in industrial systems, it often faces numerous network attacks. Therefore, intrusion detection systems are crucial for real-world IIoT protection and network security solutions to propose. However, current rule-based and machine learning-based IIoT intrusion detection systems suffer from low robustness when facing complex and dynamic network attacks. To address these challenges, this paper proposes a novel intrusion detection model named RASNet (Residual Attention Self-learning Network), based on deep reinforcement learning. The model introduces residual networks to enhance its ability to represent complex attack features effectively as well as address problems such as gradient vanishing and exploding. Additionally, the model incorporates multiple layers of attention mechanisms into convolutional neural networks, in order to improve the feature extraction capability of convolutional blocks, finally key information can be captured further. By dynamically adjusting the weights of feature maps, the model highlights crucial information related to malicious attacks, promoting accurate classification. To handle the issues of imbalanced attack samples and adaptability to various attack types, the model introduces a self-learning mechanism. This mechanism dynamically adjusts to different data distributions and attack types, enhancing the model’s adaptability and generalization capability. Through the backpropagation algorithm, the model dynamically adjusts the weights of different attack types, thereby improving the accuracy and robustness of malicious attack detection. This self-learning mechanism enables the model to accurately detect malicious attacks in the Industrial Internet of Things and enhance its robustness and reliability in various attack scenarios. Experimental evaluations are conducted on the publicly available dataset Edge-IIoTset, which contains 14 types of attacks for comparison with control group models. The results demonstrate plain improvements in classification accuracy achieved by our proposed model, with an accuracy of $94.97 \%$ for 15 -class classification. These results also validate the effectiveness and robustness of our model compared to existing machine learning-based models in terms of classification performance and accuracy.

External IDs:dblp:conf/msn/RenGMYY23