Intrusion detection models using enhanced denoising autoencoders and lightgbm classifier with improved detection performance

Sheikh, Abdul Hameed (2023) Intrusion detection models using enhanced denoising autoencoders and lightgbm classifier with improved detection performance. PhD thesis, UTAR.

Preview

PDF Download (2916Kb) | Preview

Abstract

An intrusion detection system (IDS) is a software developed to monitor network traffic for suspicious activities to secure data transmission. The conventional IDS strategies are vulnerable to distorted high dimensional network traffic. To overcome this, we proposed an IDS that combines a denoising autoencoder (DAE) and LightGBM classifier. The DAE aims to reduce the distortions in the network traffic by extracting the compressed hidden features representation. The LightGBM classifier aims to classify the samples using the histogram bins of the extracted features with larger gradients, which possibly boost the predictive capacity of the model. To eliminate the deviations in the latent structure, the DAE is enhanced. They are 1. DAE with Jacobian Gradient Norm, which minimizes the larger partial derivatives of the encoder activation 2. DAE with Iterating Thresholding Function, which minimizes the larger magnitude values of the encoder activation weight 3. DAE with Data Pairwise Similarity Weight, which groups the similar data points with strong similarity weight in the encoder activation clusters 4. DAE with Approximated Standard Normal Distribution, which approximates the latent structure to the standard normal distribution using inference strategy. To evaluate the effectiveness of the proposed models, they are experimented using various benchmark datasets. Notice that our proposed models achieve higher detection rate, which outperform the existing IDS models against all the eight commonly used datasets.

Actions (login required)