Applying Unsupervised Learning to Spot Subtle Variations in Login Patterns

Abstract

This study investigated the application of unsupervised learning techniques to detect subtle variations in user login patterns. It had the mission of raising cybersecurity standards and finding anomalies without the use of labelled information. Isolation Forest, K-means clustering, and Autoencoders algorithms were used on the secondary datasets with behavioural features of logins (time, location, and type of device). The paper has discovered that these models were good at identifying suspicious behaviour and performed better to conventional rule-based systems in respect to adaptability and accuracy. Working Case Studies of Microsoft and IBM endorsed what has been made clear with regards to the practical value of the use of unsupervised learning in the actual field. Irrespective of the elements that hindered data availability and limited the methodological approach, the study showed that unsupervised models could offer scalable, intelligent solutions to proactive threat detection in dynamic authentication conditions.