Inferring Mobile Apps from Resource Usage Patterns

Abstract

Despite many applications, mobile cloud computinginduces privacy concerns. In particular, when mobile device usersoffload the computation of a mobile app to the cloud, they may notwant the cloud service provider (CSP) to know what kind of appthey are using, since that information might be used to infer theirpersonal activities and living habits. One possible way for the CSPto learn the type of an offloaded app is to observe the resourceusage patterns of the app (e.g., CPU and memory usage), sincedifferent apps have different resource needs due to their distinctcomputation workloads. To assess this risk, this paper answers thefollowing question: Can the type of mobile app (e.g., email, webbrowsing, mobile game, etc.) used by a user be inferred from theresource usage pattern of the mobile app? We investigate theresource usage patterns of apps and whether the difference inresource usage pattern is sufficient to classify different types ofapps. Specifically, two privacy attacks under the same frameworkare proposed based on supervised learning algorithms. Then theseattacks are implemented and tested in a mobile device and in acloud computing environment. Experiments show that, when theresource usage patterns on a mobile device are used, the type ofapp can be inferred with high probabilities, when the resourceusage patterns on a cloud server are used, the type of app can beinferred with accuracy much higher than random guess.