Metaheuristic assisted hybrid deep classifiers for intrusion detection: a bigdata perspective

Abstract

The growth of social networks and cloud computing has resulted in the production of enormous amounts of data, which poses significant challenges for intrusion detection systems (IDS). Big data management in the IDS system presents several important issues, such as delayed reaction times, imbalanced datasets, reduced detection rates, and false alarm rates. To overcome those drawbacks, this work introduces a novel Intrusion Detection System from the perspective of big data handling. Here, input data is handled with the Apache Spark. In the first phase (preprocessing), improved min–max normalization is performed. Subsequently, improved correlation and flow features are extracted since the information extraction from the data is more important to determine the appropriate class differences during attack detection. Subsequently, intrusion detection is done by a hybrid model, which fuses the long short term memory and optimized convolutional neural network (CNN). Then, the optimization-assisted training algorithm called elephant adapted cat swarm optimization (EA-CSO) is proposed that tunes the optimal weights of CNN to enhance the performance of detection. Finally, the performance of the adopted model is validated over the traditional models in terms of positive, negative and other metrics, and the proposed work shows its better performance over the other models. The accuracy of detecting the intrusions using the HC + EA-CSO model at 90th LP is high around 95.029 while other conventional models obtain minimal accuracy.