Posted by Martin Pelikan, Giles Hogben, and Ulfar Erlingsson of Google’s Security and Privacy team
Mobile apps entertain and assist us, make it easy to communicate with friends and family, and provide tools ranging from maps to electronic wallets. But these apps could also seek more device information than they need to do their job, such as personal data and sensor data from components, like cameras and GPS trackers.
To protect our users and help developers navigate this complex environment, Google analyzes privacy and security signals for each app in Google Play. We then compare that app to other apps with similar features, known as functional peers. Creating peer groups allows us to calibrate our estimates of users’ expectations and set adequate boundaries of behaviors that may be considered unsafe or intrusive. This process helps detect apps that collect or send sensitive data without a clear need, and makes it easier for users to find apps that provide the right functionality and respect their privacy. For example, most coloring book apps don’t need to know a user’s precise location to function and this can be established by analyzing other coloring book apps. By contrast, mapping and navigation apps need to know a user’s location, and often require GPS sensor access.
One way to create app peer groups is to create a fixed set of categories and then assign each app into one or more categories, such as tools, productivity, and games. However, fixed categories are too coarse and inflexible to capture and track the many distinctions in the rapidly changing set of mobile apps. Manual curation and maintenance of such categories is also a tedious and error-prone task.
To address this, Google developed a machine-learning algorithm for clustering mobile apps with similar capabilities. Our approach uses deep learning of vector embeddings to identify peer groups of apps with similar functionality, using app metadata, such as text descriptions, and user metrics, such as installs. Then peer groups are used to identify anomalous, potentially harmful signals related to privacy and security, from each app’s requested permissions and its observed behaviors. The correlation between different peer groups and their security signals helps different teams at Google decide which apps to promote and determine which apps deserve a more careful look by our security and privacy experts. We also use the result to help app developers improve the privacy and security of their apps.
These techniques build upon earlier ideas, such as using peer groups to analyze privacy-related signals, deep learning for language models to make those peer groups better, and automated data analysis to draw conclusions.
Many teams across Google collaborated to create this algorithm and the surrounding process. Thanks to several, essential team members including Andrew Ahn, Vikas Arora, Hongji Bao, Jun Hong, Nwokedi Idika, Iulia Ion, Suman Jana, Daehwan Kim, Kenny Lim, Jiahui Liu, Sai Teja Peddinti, Sebastian Porst, Gowdy Rajappan, Aaron Rothman, Monir Sharif, Sooel Son, Michael Vrable, and Qiang Yan.
For more information on Google’s efforts to detect and fight potentially harmful apps (PHAs) on Android, see Google Android Security Team’s Classifications for Potentially Harmful Applications.
S. Jana, Ú. Erlingsson, I. Ion (2015). Apples and Oranges: Detecting Least-Privilege Violators with Peer Group Analysis. arXiv:1510.07308 [cs.CR].
T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, J. Dean (2013). Distributed Representations of Words and Phrases and their Compositionality. Advances in Neural Information Processing Systems 26 (NIPS 2013).
Ú. Erlingsson (2016). Data-driven software security: Models and methods. Proceedings of the 29th IEEE Computer Security Foundations Symposium (CSF’16), Lisboa, Portugal.