Intrusion detection using artificial neural networks and supervised deep learning

Abstract

One of the most challenging problems facing researchers and security experts today is protecting digital assets from the rising number of threats emerging from large networks, such as the Internet daily. Over the last decade, the frequency of cyber-attacks and the number of cybersecurity threats have risen exponentially, fuelled by rapid changing technology and a growing reliance on the Internet and information systems by society as a whole. With the ever-expanding network of data in this digital age, the importance of cybersecurity cannot be overstated. Unfortunately, however, complexities in securing information systems and computer networks against mutating threats and intrusions have given rise to a trend of detection rather than prevention. Access control, encryption and firewalls have traditionally been recommended as the first line of defence against network and system intrusions. However, given enough time and resources, even the most secure system or network may be breached by an intruder. As a result, intrusion detection systems have been proposed as a second line of defence against cyber-attacks. Intrusion detection systems, which system administrators typically use to detect security breaches in an organisation's network, could be automated and enhanced, using an artificial neural network. The main purpose of this study is to determine how a best-first search architecture optimisation algorithm could be designed and implemented to automate the construction of an accurate symmetrical autoencoder in an autoencoder-based intrusion detection model. Two autoencoder-based models were developed and trained to achieve this, using a reputable intrusion detection dataset. The initial model, which was developed manually, served as a performance baseline. The second model, which was developed by invoking the best-first search architecture optimisation algorithm, was evaluated against the baseline and other models from relevant literature. Results from the study suggest that the proposed and developed algorithm can produce symmetrical autoencoder-based intrusion detection models with accuracies comparable to, and in some cases, better than models found in the literature. Furthermore, these results indicate that the best-first search optimisation algorithm may be suitable for automating the construction of an accurate symmetrical autoencoder in an autoencoder-based intrusion detection model.