Automated Analysis Shows Apps Don’t Always Seem To Do What They Say
How a mobile app says it will collect or share a user’s personal information with third parties often appears to be inconsistent with how the app actually behaves, a new automated analysis system developed by Carnegie Mellon University has revealed.
"Overall, each app appears to exhibit a mean of 1.83 possible inconsistencies and that’s a huge number," said Norman Sadeh, professor of computer science in CMU’s Institute for Software Research. The number of discrepancies is not necessarily surprising to privacy researchers, he added, "but if you’re talking to anyone else, they’re likely to say 'My goodness!'"
Sebastian Zimmeck, a post-doctoral associate who designed and implemented this system with Sadeh, will present their findings Nov. 17-19 at the AAAI Fall Symposium on Privacy and Language Technologies in Arlington, Va.
A number of federal and state laws require mobile apps to have privacy policies, such as the Children’s Online Privacy Protection Act (COPPA) for mobile apps directed at children that collect personally identifiable information. However, CalOPPA requires privacy policies for any mobile app that collects personally identifiable information, regardless of whether it is directed at children or adults. Delaware has a similar law. Those state laws effectively serve as a minimum privacy threshold because app publishers usually don’t market state-specific apps.
Sadeh’s group is collaborating with the California Office of the Attorney General to use a customized version of its system to check for compliance with CalOPPA and to assess the effectiveness of CalOPPA and “Do Not Track” legislation.
"Just because the automated system finds a possible privacy requirement inconsistency in an app does not mean that a problem necessarily exists," Sadeh emphasized.
"That’s why a human would need to validate findings by the automated system before any enforcement or corrective action took place," Sadeh said.
Nevertheless, with substantially more than a million apps already in Google Play and the number growing by the day, such a system might help developers detect problems with their apps before they are marketed and could help make spotting violators of laws a more manageable task for regulators and privacy activists.
This approach is far faster than any human review. Two years ago, for instance, 1,200 apps were reviewed in a week’s time by the joint efforts of 26 international privacy enforcement agencies. The Carnegie Mellon system, by comparison, was able to review almost 18,000 in about 31 hours, or about 6 seconds per app.
"With a few servers, we should be able to scan all the free apps in the Google Play store every month," Sadeh said.
It would still require a lot of manpower to systematically review those apps flagged by the automated system, however. Instead, the system could be used to assign scores to apps and help regulators focus on the seemingly most egregious ones. This could result in letters or emails being sent to developers asking for clarification.
"Some discrepancies are to be expected because not all developers are sophisticated about privacy," Sadeh said.
"Whenever you’re using Google Maps," he noted, "you’re effectively sharing personal information with Google."
The National Science Foundation, the Defense Advanced Research Projects Agency and the Air Force Research Laboratory supported this work. In addition to Sadeh, the researchers included Sebastian Zimmeck, Ziqi Wang, Lieyong Zou, Florian Schaub, Shomir Wilson and Bin Liu of Carnegie Mellon’s School of Computer Science, as well as Steven M. Bellovin of Columbia University, Roger Iyengar of Washington University of St. Louis and Joel Reidenberg of the Fordham University School of Law.
Byron Spice | 412-268-9068 | email@example.com