Predicting Security of Browser Extensions Using Machine Learning

Abstract

Combining machine learning and cyber security can have great results in finding malicious web extensions and how those extensions exploit their users. In this thesis, a method for discovering extensions that are attacking their users with the help of Query Stealing attacks is proposed. With the help of machine learning, the suggested method is to build a model that with high accuracy predicts whether an extension is attacking its users or not based on static information found in both the manifest of the extensions and in the presence of predefined Keywords. The model is trained with previously gathered data that contains labels on extensions that categorise extensions as either malicious or safe. This data is suitable for a model taught using supervised machine learning. The final model achieves an F1 score of 0.9651, which indicates that the risk of an extension being misclassified is low. The model was then used to predict labels of new unlabelled extensions.