An Intrusion Detection System Based on Deep Belief Networks

The rapid growth of connected devices has led to the proliferation of novel cyber-security threats known as zero-day attacks. Traditional behaviour-based Intrusion Detection Systems (IDSs) rely on Deep Neural Networks (DNNs) to detect these attacks. The quality of the dataset used to train the DNNs plays a critical role in the detection performance, with underrepresented samples causing poor performances. In this paper, we develop and evaluate the performance of Deep Belief Networks (DBNs) on detecting cyber-attacks within a network of connected devices. The CICIDS2017 dataset was used to train and evaluate the performance of our proposed DBN approach. Several class balancing techniques were applied and evaluated. Lastly, we compare our approach against a conventional Multi-Layer Perceptron (MLP) model and the existing state-of-the-art. Our proposed DBN approach shows competitive and promising results, with significant performance improvement on the detection of attacks underrepresented in the training dataset.