Cyberspions Expose Secrets via Camera’s Movable Elements

Researchers developed methods for extracting sounds from static images taken by smartphone cameras. The moving details of the camera, such as the roller shutters of CMOS, the lenses for the optical stabilization of the image (OIS) and autofocus (AF), create sounds that are modulated in images as inconspicuous distortions.

In research work, scientists explain that smartphone cameras create a special optical-acoustic side channel, which does not require direct visibility or presence of the object in the field of view of the camera, but allows us to consider this information with high accuracy.

Having focused on the restrictions on the side channel, which is based on the “suitable mechanical path from the source of sound to the smartphone”, the researchers have removed and analyzed the leaked acoustic information, according to which several speakers can be effectively identified, their floor and even the numbers pronounced by them.

Scientists rely on machine learning to restore information from human speech transmitted by the speakers. The study was conducted from the point of view of an attacker who has a malicious application on the smartphone but no access to the microphone. Nevertheless, the threat model assumes that an attacker can shoot a video using a victim’s camera and can get samples of speech of targets in advance to use them as part of the learning process.

Using a set of data from 10,000 samples of pronouncing signal numbers, the researchers greatly improved their algorithm and set it up to perform various tasks. For experiments, scientists used the Google Pixel, Samsung Galaxy, and Apple iPhone devices.

“Our assessment using 10 smartphones in a set of data on the pronounced numbers shows 80.66%, 91.28%, and 99.67% of the recognition accuracy of 10 numbers pronounced, 20 speakers, and 2 representatives of different sexes, respectively,” the researchers said.

Researchers believe that cameras of lower