Machine learning for attacking gesture-based phone unlocking

Abstract

Since the invention of smartphones, numerous functionalities have been gradually incorporated into them over the years. Complex instruments such as gyroscopes, accelerometers and cameras have become essential to a smartphone’s functionality. However, the vast number of instruments can become points of security weaknesses that could potentially be exploited. This leaves users vulnerable to side-channel attacks, a technique that leverages on information gathered from such devices to obtain confidential information. Information can be leaked during the entry of PIN codes, such as from the change in orientation of the smartphone, or from the coordinates of the gaze of the user’s eyes. With the use of Machine Learning models, data collected from such instruments can be leveraged to infer the user’s PIN code. The initial objective of this thesis was to research the efficacy of combining different side channel information to predict keystrokes, and in turn deduce the user’s PIN code. However, a review of the data collected prior to this thesis suggests that it is of a poor quality, which requires rectification before this planned objective can be carried out. Thus, despite starting with the initial objective in mind, the data quality is not appropriate enough to justify conducting the research based on the initial objective. As such, this thesis will detail the process of the review on the work done prior, and evaluate the processes of the experiment. With revision of data collection methods, it is believed that further research can be conducted to achieve the initial research objective.