From 2008 to 2018, the use of mobile apps has grown exponentially, reaching more than 3.8 million apps in Google’s Play Store and two million apps in Apple’s App Store, followed by numerous other apps available from Amazon, Microsoft, and so on. Mobile apps cover almost all categories of users’ daily activities, including games, business, education, entertainment, health, travel, food, and more. Some apps use sensors embedded in the mobile device to collect data, such as motion, environmental, or positional information. The purpose of this study is to reveal the sensor usage patterns of these apps.

The authors present SDFDroid, software for analyzing how Chinese Android apps and Google Play apps use sensors. The app codes are disassembled into smali (Android Assembly) code where the data and operations are analyzed for their sensor types and data propagation paths. SDFDroid performs backward tracking analysis to identify sensor types used in the code, and forward tracking analysis to create a graph that shows how each identified sensor object is propagated in smali code. To understand the different apps’ usage patterns, the sensor data propagation graphs are clustered using the DBSCAN algorithm and the graph similarity measure based on the graphs’ hash between each pair of graphs. The hash value of a graph is obtained based on the hash value of all the nodes in the graph. The similarity between two graphs is computed based on the number of nodes that have the same hash values.

In total, the authors identify 98 and 35 clusters from 19914 AppChina apps and 7601 Google Play apps, respectively. Apps in the same cluster tend to either use the same third-party libraries or were developed by the same developer. Many games and display ads activate embedded sensors. Accelerometer, magnetic field, proximity, and rotational vector sensors are the most used.

The results shed light on the current usage patterns of embedded sensors in different types of apps, and on popular third-party libraries used in coding the apps. The study does not actively collect any dynamic user behaviors, which remains a challenging and important issue concerning how these sensors impact user behavior when adopting apps.