Intrusion Detection Based on Feature Selection and Transformer BiGRU

In response to the current feature redundancy in intrusion detection system traffic data and the low accuracy and slow training speed of traditional machine learning algorithms in network intrusion detection, this paper proposes an intrusion detection system based on feature selection and deep learning. Firstly, the dataset is normalized and subjected to single hot encoding processing. For redundant features that appear after single hot encoding processing, a Bayesian optimized LightGBM is used for feature selection to reduce redundant features. Then, the processed dataset is transferred to the encoder part of the transformer to capture global connections and perform preliminary feature extraction on the data. Then, DNN is used to further feature deep level features, Finally, the classification results are obtained through the softmax classifier. The experimental results on the NSL-KDD dataset show that this method has significantly improved detection accuracy in both binary and binary classification tasks.