Abstract: Because of the imbalance in existing network intrusion datasets, it is difficult for network intrusion detection systems to detect a small number of sample attacks accurately. And it is difficult to train high-dimensional data with long training time, which easily leads to poor intrusion detection accuracy. We propose an efficient intrusion detection system based on LightGBM algorithm for improving the efficiency of network intrusion detection. The system used the SMOTE up-sampling method to balance the dataset first, and performed feature dimensionality reduction on the high-dimensional dataset with the rank search algorithm to extract a subset of more important and strongly relevant features. The dataset was trained and classified with the parameter-optimized LightGBM classifier. The experimental results show that this system consumes less time and obtains higher accuracy in both binary and multi-classification experiments, which effectively improves the efficiency of intrusion detection.