About Sensibility Testbed
Ever wondered what science could achieve if any researcher can get data from other people's smartphones? Imagine that we would simply write a few lines of code, fire it up on a number of strangers’ phones, and within minutes we would know where the dead spots of our mobile data plans are. We could also have a zero-cost navigation system when no GPS or any other location services are available: we could achieve this by establishing a Bluetooth connection with a neighboring device and get the location data from it. If we constantly monitor accelerometer data on mobile devices, we can detect vibrations within the frequency and intensity range of seismic waves, and assist distributed earthquake detection. These all sound fantastic, except that who would let us get data off their devices? Our friends and family would probably trust us. Other people? Not so much.
The privacy and security challenges on mobile devices have increased dramatically over the years. While having apps post tweets to user’s Twitter account without asking for permission is seriously off-putting, hacked apps that let criminals break into individual’s bank account are detrimental. Running code to collect data from smartphones is much more complex than it sounds. We not only need to ensure the security of a device so that the code that does the data collection cannot damage or hack into the device. We also must protect the privacy of a device owner so that the code cannot eavesdrop phone conversations, steal passwords, etc.
Sensibility Testbed is a smartphone testbed that allows researchers to run code and perform measurements on others’ smartphones for research purposes. It ensures the security of user-owned devices, and the privacy of user-generated data. The usage model of Sensibility Testbed is unique in that it manages how device owners make their devices accessible to different research communities without putting their devices at risk. Meanwhile, it offers technical resources that allow researchers to collect data from remote mobile devices without impairing the device owner’s privacy. As an added bonus, different research groups can pick, choose, share and reuse each others' user base.
Sensibility Testbed: Tutorial and Demo
Download To donate resources to the Sensibility Testbed, please download our app from Google Play.
Setup The setup and instructions of how to run code on the Sensibility Testbed, and the API to access currently supported sensors can be found at project demo page.
Local debug If you would like to debug Sensibility code locally on your phone, here is a way to do it.
Sensibility Testbed Overview
Researchers run experiments in Sensibility Testbed by writing code for a restricted, Python-based sandbox. This is the same security-reviewed Repy (Restricted Python) sandbox used in the Seattle testbed . This sandbox has been deployed on the Seattle testbed over the last 6 years. Our experience has shown that the risk of it being faulty is very low. The Repy sandbox is restricted in that its API limits what a sandboxed program can do: reading from and writing to the file system can only occur in a per-experiment directory; sending and receiving data via the network interface cannot exceed a configured rate; CPU, memory and battery consumption cannot exceed a limit, etc. Therefore, the sandbox isolates the program from the rest of the device. More importantly, the sandbox allows us to interject code to implement privacy policies and control what happens with the data gathered on the device.
Why Sensibility Testbed
First, in our testbed, device owners participate as volunteers, and researchers request these devices through our server. This server, which is called a clearinghouse, mediates remote device access, but does not store any personal data. As a result, Sensibility Testbed relieves researchers from recruiting participants and allows different groups to share all the devices participated in the testbed. Additionally, Sensibility Testbed does not require researchers to write full-fledged Android apps to perform experiments. Instead, it provides an easy-to-use Python-like programming interface. Last but not the least, using Sensibility Testbed, device owners do not need to trust the researchers who run code on their devices. As you will see later in this article, Sensibility Testbed provides a secure, sandboxed environment for anyone to run experiment code on Android devices. This can effectively prevent potential security and privacy breaches. If you are interested in how to setup a clearinghouse, please refer to this page.
How Sensibility Testbed Works
In Sensibility Testbed, there are three categories of interacting parties: mobile devices owned by ordinary people, with our app installed; a clearinghouse server that discovers and configures participating devices; and researchers wanting to run experiments on mobile devices (see Figure 1). These three parties interact as follows. Mobile devices provide resources and data for researchers to use in their experiments. As mentioned above, researchers’ code runs in a sandbox on a remote device that isolates the code from the rest of the device host system. Meanwhile, our clearinghouse server helps researchers acquire and manage devices, and enforces policies specified by the researcher’s Institutional Review Board (IRB), thus protecting device owners' personal information. Finally, researchers use their local machines to initiate and control experiments in Sensibility Testbed. They use an experiment tool (ET) to deploy and run experiments in sandboxes on remote devices that are acquired through the clearinghouse.
Figure 1: Sensibility Testbed architecture.
In the following, let us suppose a researcher, Bob, runs an experiment on a device owner, Alice's smartphone. To understand how policies are implemented, we need to start with the Sensibility Testbed app. The app on Alice’s device contains a native Android portion, and our Repy sandbox. When Alice starts the app, the native code initializes a Python interpreter, launches the Repy sandbox, and starts the communication between the device and our clearinghouse. The sandbox’s restricted, secure API provides calls to file system, networking, threading functions, and so on. Therefore, Bob’s code can read files, send data through the network, etc., from Alice’s device. However, the original Repy sandbox does not include calls specific to mobile devices, such as GPS location, WiFi network, Bluetooth, accelerometer, cellular network, etc.
To obtain smartphone-specific data, we first implemented our sensor API using native code in the Sensibility Testbed Android app. The Repy sandbox then uses RPC to invoke the corresponding Android code, and returns the data from native code to a sandboxed program. The Repy sandbox thus defines a set of sensor API as higher level calls, such as get_location(), get_wifi(), get_accelerometer(), etc. As such, the original Repy interface and the added sensor API together provide the complete "OS level" sandbox kernel on a mobile device, as shown in Figure 2.
Figure 1: Sensibility Testbed blur policies.
Finally, this sandbox kernel determines how policies are implemented by affecting API calls. It can interpose on a call and modify the data returned, or control the frequency a call can be made over time. As mentioned above, Bob provided his IRB policies through our clearinghouse. So before Bob runs his experiment, the clearinghouse instructs the sandbox on Alice’s device to restrict sensor access in accordance with these IRB policies. Using the get_location()call as an example, when Bob's code requests location data from Alice’s device, the Repy sandbox first invokes the location-related Android code. As the location data is returned, Bob’s IRB policy indicates that the returned location coordinates should be blurred to the nearest city to Alice’s device, instead of her actual location. As a result, the sandbox returns an approximate location to Bob’s program. Furthermore, as Bob’s IRB policy disallows collecting information about cell tower IDs, the access to cell IDs is blocked entirely on Alice’s device. Similarly, other information like WiFi SSID can be blurred to a hashed string, the frequency to access an accelerometer can be restricted to prevent inferring passwords from the movement and tilt of the device, and so on. As shown in Figure 2, different policies can be stacked together as a set of filters for different sensors, before a sandboxed program can access the sensor data.
This page lists all the currently supported sensors, and how to use our API to access these sensors.
Related Publications and Websites
 J. Cappos, L. Wang, R. Weiss, Y. Yang, and Y. Zhuang. BlurSense: Dynamic fine-grained access control for smartphone privacy. Sensors Applications Symposium (SAS), 2014. (pdf)
 A. Rafetseder, F. Metzger, L. Puhringer, K. Tutschku, Y. Zhuang, and J. Cappos. Sensorium–a generic sensor framework. Praxis der Informationsverarbeitung und Kommunikation, 36(1):46–46, 2013. (pdf)
 Open3GMap. https://o3gm.cs.univie.ac.at/
 J. Cappos, A. Dadgar, J. Rasley, J. Samuel, I. Beschastnikh, C. Barsan, A. Krishnamurthy, T. Anderson. Retaining Sandbox Containment Despite Bugs in Privileged Memory-Safe Code. ACM Conference on Computer and Communications Security (CCS'10, pdf)
Here is the blog which reports some of our latest progress.
Appropriate Use Panel
For information how to obtain our data or share sensor data with us, please visit Sensibility Testbed Appropriate Data Use Panel page.
- arch.jpg (147.7 KB) - added by yyzhuang 10 months ago.
- arch-small.jpg (59.7 KB) - added by yyzhuang 10 months ago.
- blur.jpg (224.1 KB) - added by yyzhuang 10 months ago.
- blur-small.jpg (92.8 KB) - added by yyzhuang 10 months ago.
- wifilock.png (1.8 KB) - added by yyzhuang 10 months ago.
- sl4a.png (1.5 KB) - added by yyzhuang 10 months ago.
Download all attachments as: .zip