Led by Dr. Shivam Bhasin, the research team "trained" the system by getting three people to each enter a random set of 70 four-digit PINs on Android smartphones. On each phone a custom app was installed, that collected data from the accelerometer, gyroscope, magnetometer, proximity sensor, barometer, and ambient light sensor.

A deep learning algorithm analyzed the data, matching up specific sensor readings with specific numbers being pressed on the onscreen keypad. "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," explains Bhasin. "Likewise, pressing 1 with your right thumb will block more light than if you pressed 9."

When the system was subsequently tasked with guessing four-digit PINs based only on the sensor reactions, it was 99.5 percent accurate (within no more than three tries) at unlocking a phone that used one of the 50 most common PINs.