Improvisation of Predictive and Prescriptive Models Using Visual Analytics and Neural Networks

Abstract

The intrusion detection system is an area of research in cyber and data security. Accurate intrusion detection is a big challenge for conventional methods such as data mining, statistical evaluation and Artificial Neural Network. Machine learning algorithms have great potential for intrusion detection. The machine learning algorithms provide flexible classification and clustering algorithms for detecting intrusion. This thesis focuses on the novel method for intrusion detection using deep learning. Deep learning is an extension of machine learning and improves the detection accuracy of IDS. The proposed algorithm cascaded three levels of Convolutional Neural Network (CNN). The cascading levels of CNN increase the feature mapping space of propagation forwarded connected network. newlineA cutting-edge intrusion detection system is proposed that is based on ensemble classification methods. In order to choose the best subset based on the correlation between characteristics, an EM method is suggested. The classification model is then built using the ensemble classifier based on SVM, DT, and KNN. Finally, 10 fold cross-validation over three intrusion detection datasets is used to assess the proposed IDS. newlineThe acquired findings for the NSL-KDD dataset show accuracy of 99.52 percent and 0.15 percent False Acceptance Ratio (FAR) with a subset consisting of just 8 features, while the experimental results for the NSL-KDD dataset show accuracy of classification equal to 99.81 percent and 0.08 percent FAR with a subset of 10 features. Surprisingly, on the subset of 13 features for the CIC-IDS2017 dataset, our model obtains the greatest accuracy of 99.89 percent and DR of 99.9 percent. The comparison with no feature selection technique then shows promising results on a number of measures, and it should be highlighted that our solution drastically cuts the SVM on the CIC-IDS2017 dataset from 97.94 to 98.42. Our approach also outperforms related feature selection with respect to ML, efficiency, and FAR limitation at modest levels. Addit