Cyber attacks are a major threat to countries, large organisations, and small businesses alike, with serious political, legal, and especially economic consequences. With these cyber attacks growing exponentially, we are faced with a major challenge which is to develop sophisticated and effective techniques to detect intrusions in the information systems of organisations in order to increase data and services protect. The majority of intrusion detection systems use signature-based approaches and supervised learning methods that depend on labelled training data. Generating this training data is usually a costly endeavour. In this study, we use autoencoders in unsupervised machine learning methods to improve intrusion detection system by performing binary traffic classification. The models were trained on four different types of autoencoders and with two datasets: CSE-CIC-IDS2018 and a dataset we generated with our home made scripts. The most relevant features for each dataset were selected to improve model generalisation, reduce complexity and improve performance. The performance of our models is evaluated in terms of accuracy and false positive rate. The error threshold of the used autoencoders is fixed with accordance to a specific objective. According to the obtained results, out of the four types of autoencoders we have studied, denoising autoencoder provides the best performance.