Feature selection using autoencoders in neural network based intrusion detection system

Abstract

An Intrusion Detection System (IDS) can identify attacks by analysing data traffic patterns within a network. Given the substantial volume of data handled by the IDS, there is a necessity to conduct feature extraction to reduce the computational burden of processing raw data. Feature extraction involves converting features into a lower dimension to expedite the learning process and enhance accuracy. This study focuses on automatic feature extraction utilizing a simple autoencoder and Artificial Neural Network for classifying (Both Binar classification and Multiclass Classification) attacks detected by the IDS. Various activation functions and loss functions are employed to access the extent to which feature extraction can enhance accuracy. The effectiveness of the detection mechanisms post feature extraction is evaluated using datasets such as KDD Cup '99. In the proposed model, the ReLU activation function and cross-entropy loss function as hyperparameters for the autoencoder yield the highest accuracy compared to other functions.