Cost-sensitive stacked long short-term memory with an evolutionary framework for minority class detection

Abstract

‘Minority attack detection’ is a matter of great concern while designing a secure network and safeguarding it against cyber criminals who attempt to breach its defenses, intrude into its systems, and gain unauthorized access to sensitive information. An adequate intrusion detection system is demanded to counter such threats involving class imbalance and high dimensionality in network traffic. In this context, an artificial intelligence-based multiclass framework MCGAL is presented, which leverages the evolutionary technique of genetic algorithm to achieve the finest feature selection and stacked long short-term memory (LSTM) for traffic classification. The proposed approach utilizes cost-sensitive LSTM by assigning greater significance to minority or rarely occurring classes, thereby addressing the challenge of misclassification in the presence of imbalanced data in a real web attack subset of the CSE-CIC-IDS2018 dataset. A comprehensive performance analysis is performed to evaluate our proposed scheme concerning various class metrics using weighted and macro averages, which reveals a classification accuracy of 100 %. An achieved weighted average of 1.00 is then compared with baseline classifiers like decision trees, AdaBoost, and support vector machines utilizing both linear and radial kernels. The results of MCGAL are also assessed with recently reported techniques on the same dataset demonstrating remarkable improvements in all metrics such as precision, recall, and F-score of minority class detection. A pair-wise comparison and Kruskal-Wallis test at a 95 % significance level further illustrate the statistical significance of the proposed scheme for minority class detection.