Inference of User Demographics from Android Smartphone Sensors
While features and functionality shipped with smartphones facilitate many aspects of our everyday life, they also pose real privacy concerns. Many sensors on smartphones allow application developers to access them without explicitly asking for the user’s permission, as the sensors are considered benign and free of any sensitive information. However, previous research has shown that you can infer user activities and movements from accelerometer readings, and even recognize individuals based on their voices using the gyroscope. Can these sensors be used to infer other sensitive information about the owners of the devices?
Research goal: Using a previously collected dataset, train a classifier to infer user demographics (including age, gender, income) from sensor readings on their smartphones. As an additional step, we can also use our Android testbed to infer how frequently apps collect data from these sensors (and see if they notify the user somehow).
How you can engage: The existing classifier is performing at just above chance level. There are multiple steps that could be improved, including sensor signal preprocessing, feature extraction and classifier design (for instance, features could be learned using neural nets as opposed to using hand-crafted features).