TouchTrack is first of its kind that goes beyond the tracking of virtual identities and demonstrates how user behavioral biometrics, in implicit mood, can identify the real person behind the mobile device(s), thus tracking the physical identification. It quantifies the uniqueness of touch-based behavioral biometrics gestures, which are performed by a user on a mobile device.
TouchTrack allows users to perform various gestures such as swipe up, down, left, right, tap, type, write a word etc through interesting games. After performing gestures, users will be presented with three different types of uniqueness results, i.e. entropy of individual features, aggregated features (per gesture), and aggregated gestures (overall). These results are calculated through an analytical framework which uses established information theoretic metrics for measuring the level of information contained in touch features, such as information surprisal and entropy, and to quantify the level of uniqueness of behavioural biometrics. The framework considers individual features as well as, possibly dependent, aggregated features as carriers of information about users.
A high-level architecture of an overall process is shown below:
Data Collection and Preprocessing
We have collected a large number of gesture samples for feature characterization and to be used as ground truth for the estimation of uniqueness of features. The dataset with initial numbers of samples serves two purposes: to identify relevant features to be considered for fingerprinting users and to train our analytical framework to evaluate the uniqueness of each feature. The initial dataset is also used for the purpose of studying dependence between features.Our initial version of the mobile “touch-analytics” app instructed the user to perform specific gestures (such as left/right swipes and taps), and “writing” and typing a few words on the touchpad. During this phase, we collected an exhaustive list of features. We will then create separate profiles for each user (identified by an ID), which we will store at our server.
Analytical Framework: an Information Surprisal Probabilistic Model
Our analytical framework is constructed based on well established information theoretic metrics to capture the amount of information carried by each of the considered features. We have used information surprisal as a measure of the amount of information about the individual being fingerprinted (tracked), where each bit of information reduces the number of possibilities of identification of the individual amongst a population in half (entropy and Information surprisal are deeply related as entropy can be seen as the expected value of information surprisal). Perhaps a more challenging task is to design the probabilistic framework to estimate the information surprisal of a combination of features that are collected from a single individual. Features are very unlikely to be independent and we have taken into account dependence of features when designing our analytical framework. We have validated feature dependence using empirical evidence extracted from our dataset.
We envision several possible uses of TouchTrack. Possible beneficiaries are: (i) end users, who may use the “touch-analytics” app to assess the uniqueness of their interactions with
- End users, who may use the “touch-analytics” app to assess the uniqueness of their interactions with touch screen devices and take appropriate decision in case they wish to stay anonymous. The community as a whole will benefit by increased awareness of this mode of tracking;
- Device providers, who can devise ways to prevent sensors from leaking identifiable information (e.g., by employing random noise);
- Service/application providers, who can use our results to modify the design of mobile applications such that only “need to know” information is obtained, instead of revealing all gesture-related data