Novel Transformation Operators for Twin-Autoencoder Applied to Intrusion Detection

Intrusion detection systems (IDSs) are important components in protecting information systems from cyberattacks. Machine Learning (ML) has played a key role in the effectiveness of IDSs by learning the pattern of network traffic in the training data and using the learnt knowledge to predict the new incoming data. Among machine learning models, Auto encoders are popular used due to their ability to transform the raw data into a new representation that facilitates the detection of cyberattacks. Recently, Twin-Autoencoder were proposed to leverage the previous AE models by deterministically transforms the latent representation into a separable representation and then learn to reconstruct the separable representation at the output. However, the transformation operator in TAE is complex and ineffective in some scenarios. In this paper, we propose two new transformation operators to improve the effectiveness of TAE. The proposed solutions are evaluated on three benchmarking IDS datasets and the experimental results show the superior performance of the new transformation operators compared to TAE and the standard AE model.