An Optimized Auto-Encoder based Approach for Detecting Zero-Day Cyber-Attacks in Computer Network

Machine Learning and Deep Learning have been applied in Cybersecurity for more than a decade, such as cyber-attack detection, intrusion detection, network traffic classification, and much more. However, detection of Zero-day cyber-attacks is the utmost priority of the security administrator. Zero-day cyber-attacks try to exploit the system's vulnerability that remains unclosed until the exploit has occurred. The solution proposed in this work is based on an Intrusion Detection System that can detect Zero-day and unknown cyber-attacks. We used the autoencoder to build an intelligent intrusion detection model. The novelty of the proposed work is to show that how threshold plays a crucial role in the detection of Zero-day cyber-attacks with good recall. Also, choosing a single threshold for one type of attack might not work effectively for other unseen cyber-attacks. Hence we have evaluated accuracy separately for each attack with different thresholds to show its significance. We have used CICIDS2017, the latest dataset, for evaluation purposes. The model shows an excellent result in terms of accuracy or recall both separately and overall. The overall accuracy of the optimized version of autoencoder (OPT_AE) is 99.29 % on the CICIDS2017 dataset.