Towards Ubiquitous Privacy Decision Support: Machine Prediction of Privacy Decisions in IoT

We present a mechanism to predict privacy decisions of users in Internet of Things (IoT) environments, through data mining and machine learning techniques. To construct predictive models, we tested several different machine learning models, combinations of features, and model training strategies on human behavioral data collected from an experience-sampling study. Experimental results showed that a machine learning model called linear model and deep neural networks (LMDNN) outperforms conventional methods for predicting users’ privacy decisions for various IoT services. We also found that a feature vector, composed of both contextual parameters and privacy segment information, provides LMDNN models with the best predictive performance. Lastly, we proposed a novel approach called one-size-fits-segment modeling, which provides a common predictive model to a segment of users who share a similar notion of privacy. We confirmed that one-size-fits-segment modeling outperforms previous approaches, namely individual and one-size-fits-all modeling. From a user perspective, our prediction mechanism takes contextual factors embedded in IoT services into account and only utilizes a small amount of information polled from the users. It is therefore less burdensome and privacy-invasive than the other mechanisms. We also discuss practical implications for building predictive models that make privacy decisions on behalf of users in IoT.