Improved convolution neural network integrating attention based deep sparse auto encoder for network intrusion detection

Network intrusion detection (NID) is seen as a pivotal technology in the network security which can detect malicious threats occurring in the network and lend stabilized services for expanding the network environments. However, Network-based Intrusion Detection Systems (NIDSs) are not sensitive to infrequent intrusions features and tend to have high misjudgment rate on imbalanced datasets, which lead to obvious defects for detecting minority classes of intrusion. Therefore, a novel neoteric NID methodology predicated upon an optimized convolutional neural network (CNN) integrating Attention based Deep Sparse Auto Encoder (ADSAE) (ADSAE-CNN) is put forward. The data expanding method based on the ADSAE model integrates attention mechanisms with deep stacked autoencoders to expand intrusion records of minority classes in data preprocessing, so as to balance the data distribution of intrusion detection datasets, improve the sensitivity of the detection model to the intrusion of few categories, and enhance the monitoring of the intrusion by the system. Meanwhile, the ADSAE can encode and transform the intrusion data to ameliorate the feature extraction capability of convolutional layers of the proposed ADSAE-CNN for detecting and classifying different intrusions classes. Finally, the ADSAE-CNN methodology is devoted to the network intrusion detection of two experiments on UNSW-NB15 and CSE-CIC-IDS2018 datasets and achieves the total precision of detection 89.1% on UNSW-NB15 and 94.20% on CSE-CIC-IDS2018, which can lead to considerably elevate the detection rate of minority intrusions and means significant effectiveness on multi-class network intrusion detections.