Machine learning for attacking gesture-based phone unlocking

Abstract

With the rapid evolution of technology, the use of smart phones in our everyday lives has increased at a tremendous rate. However, our rising reliance on smartphones for our day-to-day activities exposes us to some security concerns. Nowadays, smartphones are equipped with a variety of devices, such as cameras, sensors like accelerometers and gyroscopes. Various applications that we use have access to these devices without our knowledge. A side-channel attack uses information acquired from the system’s implementation via these devices to obtain sensitive information. Through the use of these instruments or side channels, an attacker could potentially recover information, such as PIN codes, which would give them unauthorised access to the victim's devices. The side channels could leak information like the variation of light, orientation of phone and even coordinates of the eyes and head when the victims are typing their PIN code. According to the current state-of-the-art research, machine learning and deep learning techniques are applied on either the sensor or camera data to retrieve PIN codes with a good accuracy. The objective of this project is to process, analyse and combine the various types of side-channel information to train a model to retrieve PIN codes with a higher accuracy. For this purpose, we create a dataset approximately with 9.8K unique points. The best results were achieved using a Decision Tree Classifier model. The accuracy to predict whether digit ‘0’ is present in the PIN code or not was 55% while the accuracy for digit ‘1’ was 54%. Based on the literature analysis, the results obtained are conclusive and are of good standard. Future works will focus on using these results to retrieve the entire PIN code sequence.