Instruments in smartphones such as the accelerometer, gyroscope, and proximity sensors represent a potential security vulnerability, according to researchers from Nanyang Technological University, Singapore (NTU Singapore), whose research was published in the open-access Cryptology ePrint Archive on Dec. 6.
Using a combination of information gathered from six different sensors found in smartphones and state-of-the-art machine learning and deep learning algorithms, the researchers succeeded in unlocking Android smartphones with a 99.5 percent accuracy within only three tries, when tackling a phone that had one of the 50 most common PIN numbers.
The previous best phone-cracking success rate was 74 percent for the 50 most common pin numbers, but NTU’s technique can be used to guess all 10,000 possible combinations of four-digit PINs.
Led by Dr. Shivam Bhasin, NTU Senior Research Scientist at the Temasek Laboratories at NTU, researchers used sensors in a smartphone to model which number had been pressed by its users, based on how the phone was tilted and how much light is blocked by the thumb or fingers.
The researchers believe their work highlights a significant flaw in smartphone security, as using the sensors within the phones require no permissions to be given by the phone user and are openly available for all apps to access.
The team of researchers took Android phones and installed a custom application which collected data from six sensors: accelerometer, gyroscope, magnetometer, proximity sensor, barometer, and ambient light sensor.
“When you hold your phone and key in the PIN, the way the phone moves when you press 1, 5, or 9, is very different. Likewise, pressing 1 with your right thumb will block more light than if you pressed 9,” explains Bhasin, who spent 10 months with his colleagues, David Berend and Dr. Bernhard Jungk, on the project.
The classification algorithm was trained with data collected from three people, who each entered a random set of 70 four-digit pin numbers on a phone. At the same time, it recorded the relevant sensor reactions.
Known as deep learning, the classification algorithm was able to give different weightings of importance to each of the sensors, depending on how sensitive each was to different numbers being pressed. This helps eliminate factors which it judges to be less important and increases the success rate for PIN retrieval.
Although each individual enters the security PIN on their phone differently, the scientists showed that as data from more people is fed to the algorithm over time, success rates improved.
So while a malicious application may not be able to correctly guess a PIN immediately after installation, using machine learning, it could collect data from thousands of users over time from each of their phones to learn their PIN entry pattern and then launch an attack later when the success rate is much higher.
Professor Gan Chee Lip, Director of the Temasek Laboratories at NTU, says this study shows how devices with seemingly strong security can be attacked using a side-channel, as sensor data could be diverted by malicious applications to spy on user behavior and help to access PIN and password information, and more.
“Along with the potential for leaking passwords, we are concerned that access to phone sensor information could reveal far too much about a user’s behavior. This has significant privacy implications that both individuals and enterprises should pay urgent attention to,” says Gan.
Bhasin says it would be advisable for mobile operating systems to restrict access to these six sensors in future, so that users can actively choose to give permissions only to trusted apps that need them.
To keep mobile devices secure, Bhasin advises users to have PINs with more than four digits, coupled with other authentication methods like one-time passwords, two-factor authentications, and fingerprint or facial recognition.
Source: Nanyang Technological University