Novel Feature Extraction Methods for Authentication via Mouse Dynamics with Semi-Supervised Learning

Abstract

Exploring novel security layers in academia and industry is always a concern due to existence paradigm between security experts and malignant activities. Adding a widely applicable security layer into existing ones in term of authentication can be achieved by behavioural biometrics. A great candidate is users’ mouse usage behaviours which is less intrusive and inimitable in terms of other behavioural biometrics and contains less sensitive data. This work will present a performance evaluation study on semi-supervised methods for fast authentication in case of data deficiency. The Balabit data set is used in experiments which provide raw mouse usage data of 10 users in split sessions. The raw data in the data set is segmented according to the actions and novel feature extraction methods are proposed. A different classifier for each action type is trained only with the target user data. Parameter selection was performed with a small amount of abnormal data represented by other users in the data set. An inlier score is generated for successive actions and the probability of being an anomaly for each individual test session is calculated. Overall, the model performance is evaluated in terms of Area Under Curve (AUC) and Equal Error Rate (EER). Comprehensive experiments show that our model’s performance is comparable with the models based on supervised methods.