Authentication of Swipe Gestures in Smartphones using Sensor Data

Abstract: Smartphones are an increasingly important part of our lives. As services that treat sensitive information are found on most modern phones, there is a growing need for security. The common approaches of entering a PIN-code or some pattern when accessing an application on the phone is not necessarily sufficient, as they are vulnerable to different types of attacks. The embedded sensors of modern smartphones enable new types of classification based on behavioural patterns. The built in accelerometer and gyroscope measures the acceleration and rotation of the phone in three different directions, respectively. This work investigates if the data collected from these sensors during a swipe-gesture contains any information about the individual conducting it, such that the individual can be distinguished from others. The data used, as opposed to much of the work that has been conducted in the field, is collected from individuals in their everyday-life. This proves to be a more challenging problem and a method, the Modified Hausdorff distance, performing well when touch-gestures are collected in a controlled environment is refuted. Furthermore, it is shown that there are differences between the different phone-models and operative systems for many features. In an attempt to capture individual information, spectrograms of the signals from the sensors are computed. Different ways of computing these are discussed and tested. A convolutional neural network is used to classify the images in a supervised setting, achieving a test accuracy of 36 % for 42 Android-users and 55 % for 18 iOS-users. Also, using autoencoders for feature representation of the spectrograms in an unsupervised setting is tested. When using kernal density estimation, equal error rates of about 40 is achieved.