A Novel Intrusion Detection System Based on Random Oversampling and Deep Neural Network

Abstract

This study proposes a new anomaly-based intrusion detection technique (IDT) that can successfully detect all types of attacks with comparatively higher accuracy. Since imbalanced datasets have the potential to compromise the performance of any classifier, the proposed study emphasizes on pre-processing of the imbalanced data using a random over-sampling algorithm. The sampled data are further subjected to a deep neural network, which is trained by different high-end optimizers, like; as Adam, Adadelta, Adagrad, Adamax, Nadam, RMSprop, and SGD to verify the efficiency of the model on testing with each optimizer. The IDT model has been evaluated using the famed KDD Cup 1999 dataset. It is observed that the proposed IDT outperforms many other state-of-the-art IDTS modelled on the same dataset in terms of accuracy to detect all attack types, namely dos, normal, probe, r2l, and u2r attack types.