An Efficient Intrusion Detection System Using Unsupervised Learning AutoEncoder

As attacks on the network environment are rapidly becoming more sophisticated and intelligent in recent years, the limitations of the existing signature-based intrusion detection system are becoming more evident. For new attacks such as Advanced Persistent Threats (APT), the signature pattern has a problem of poor generalisation performance. Research on intrusion detection systems based on machine learning is being actively conducted to solve this problem. However, in the actual network environment, the attack sample is collected less than the normal sample, so it suffers a class imbalance problem. When a supervised learning-based anomaly detection model is trained with these data, the results are biased toward normal samples. In this paper, AutoEncoder (AE) is used to perform single-class anomaly detection to solve this imbalance problem. The experimental evaluation was conducted using the CIC-IDS2017 dataset, and the performance of the proposed method was compared with supervised models to evaluate the performance.