Generative Adversarial Network and Bayesian Optimization in Multi-class Support Vector Machine for Intrusion Detection System

Abstract

Network Intrusion detection performances are highly affected by imbalance data problems due to the presence of less number of attack information in the dataset. Deep learning models are applied in existing methods to improve the efficiency that has limitations of overfitting problems. The Generative Adversarial Network (GAN) – Bayesian optimization Multi-class Support Vector Machine (BMSVM) is proposed to overcome imbalance and overfitting problems in intrusion detection systems. The Min-Max Normalization method is applied to normalize the input data to reduce the differences in features. GAN model is applied to generate minority class to balance the data instances to train the model. The proposed GAN-BMSVM model is compared with the classical sampling method, optimization, and classifier in the intrusion detection model in terms of Accuracy, Detection Rate (DR), and False Alarm Rate (FAR). The classical sampling methods are Near-miss, SMOTE and Autoencoder; traditional classifiers are KNN, RF, SVM, DNN and LSTM, and classical optimizations are PSO and WOA. The existing researches such as HCRNN, HLD, DONN, FL-NIDS and CNN-LSTM are used to evaluate the efficiency of GAN-BMSVM model. The GAN-BMSVM model has achieved 99.58% and 85.38 % accuracy for NSL-KDD and UNSW-NB15 dataset respectively, which is higher than the existing CNN-LSTM model