Comparative Analysis of Feed-Forward and RNN Models for Intrusion Detection in Data Network Security with UNSW-NB15 Dataset

The paper presents an improved method of network intrusion detection using machine learning based approach. The study trained four types of deep neural network models, including two feed-forward dense models and two models that utilized long-short term memory layers. The study compares the effectiveness of models predictions based on the newly developed dataset UNSW-NB15. A suitable part of the dataset was selected after evaluating its integrity, and the selected data was preprocessed and formatted for neural network training and evaluation. Two models consisting of dense layers were developed, trained, and fine-tuned to optimize their accuracy. The LSTM models were also implemented and fine-tuned accordingly. The accuracy of models' predictions was evaluated using a portion of the dataset, the validation set, which the model had not seen in training. Without weight balancing, the dense-layer-based model attained a validation accuracy of 78.94% while the LSTM-based model achieved 76.84%. With weight balancing, the accuracy improved to 79.34% for the dense-layer model and 79.21% for the LSTM model. Significant output differences were found when analysing prediction correctness of confusion matrices.