基于特征重构与MIR-BiLSTM的入侵检测方法OA

In view of the low detection accuracy caused by high computational complexity and poor adaptability to unbalanced data in current deep learning based intrusion detection models,this paper designs a hybrid detection model based on feature reconstruction and lightweight MIR-BiLSTM.Firstly,the SMOTE-ENN algorithm combining oversampling and undersampling is used to eliminate the category imbalance.Secondly,the improved stacked auto encoder(SAE)is designed to downscale and enhance the input features,and the low-dimensional robust feature representation is extracted by unsupervised pre-training.Next,the MIR-BiLSTM architecture including deformable convolutional LBL module is constructed,which achieves efficient feature extraction by combining multi-branch spatial modeling and bi-directional timing analysis.Finally,a staged training strategy is used to optimize the feature-task fitness.Experiments show that the proposed model requires a parameter count of only 2.76 MB(51.3%lower than MobileViT)and achieves an accuracy of 98.15%and an F1-score of 97.35%on the test set,which outperforms the existing lightweight intrusion detection models.To sum up,it provides an effective solution for edge computing environments.