A Novel Feature Prioritization Algorithm in Multi-Valued Neutrosophic ConvLSTM for Improving Intrusion Detection

Abstract

Cyber security is crucial due to the increasing impact of networks on modern life. As network sizes and demands increase, attackers can create new types of attacks. Therefore, developing effective Instruction Detection Systems (IDS) is essential. Recently, Various Deep Learning (DL) techniques have been proposed for automatic IDS and abnormal behavior identification of networks. Amongst, Multi-Valued Neutrosophic Convolutional Long Short Term Memory (MVN-ConvLSTM) model is developed for efficient IDS which retrieves various deep features to generate predictive uncertainty estimation. But, this model lacks the training stage accuracy in case of the large number of irrelevant attributes. To solve this, Feature prioritized MVN-ConvLSTM (FMVN-ConvLSTM) is developed in this paper to prioritize the features and reduces the complexity for efficient IDS. In this method, in-model technique is developed as the Feature Ranking (FR) algorithm allowing for accurate interpretation without compromising performance. This model utilizes the inner dynamics of ConvLSTM to determine every attribute's significance independently of any additional method. It ensures that each feature's importance will be caused by the ways in which it interacts with other attributes addressing the interpretability tradeoff.  Also, norm is utilized instead of the  norm due to the impact of and  regularizations based on the FR importance in the model. The model's configuration ensures sparse solutions resulting insignificant attributes with near zero values using  function. The constructed feature prioritizing algorithm will be added as layer in ConvLSTM model. Moreover, ConvLSTM with weighted softmax is employed to address the imbalanced intrusion classification problem, enabling better feature prioritization and preventing misclassification for IDS. Finally, the test results show that the FMVN-ConvLSTM achieves accuracy value of 0.96%, 0.95% and 0.95% compared to the other existing models on three different datasets like CIC-IDS2018 WSN-DS and UNSW-NB15.