Predicting User Activities and Device Interactions Using Adversarial Sensor Data: A Machine Learning Approach

Smart device has become a powerful tool for people of all ages. However, the elderly population is generally less aware of how such technology provides support and benefits through the variety of its applications: healthcare care, social interaction, entertainment, and more. Such a lack of awareness poses a vulnerability amongst the elderly user group when faced with an adversary attack, as granting sensor data access to the application operator may result in compromising the user's privacy. This paper studies whether user activities and device interaction can be compromised (or predicted) by adversary sensor data. Precisely, we collect built-in sensor data from the accelerometer, gyroscope, and touchscreen sensor, and seek to make predictions on the routine activities of the user. We will perform supervised learning on the collected dataset using two textbook classifiers, namely, the Decision Tree (DT) and the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$K$</tex>-Nearest Neighbours (KNN). Our experiment shows that these simple classifiers can provide reasonable prediction accuracy, indicating the presence of the leak of side information from adversary sensor data. Specifically, the prescribed classifiers achieve a test accuracy of ~ 80% when being trained over the raw data feature.