Using an off-the-shelf automotive radar sensor and a novel processing approach, Pennsylvania State University researchers demonstrated they could detect the vibrations of a cell phone’s earpiece and decipher what the person on the other side of the call was saying with up to 83% accuracy.
The demonstration, available in the 2022 IEEE Symposium on Security and Privacy (SP), reveals a significant security concern, according to Mahanth Gowda, assistant professor of computer science and engineering, and doctoral candidate Suryoday Basak.
“As technology becomes more reliable and robust over time, the misuse of such sensing technologies by adversaries becomes probable,” Basak said. “Our demonstration of this kind of exploitation contributes to the pool of scientific literature that broadly says, ‘Hey – automotive radars can be used to eavesdrop audio. We need to do something about this.’ ”
The radar operates in the millimetre-wave (mmWave) spectrum, specifically in the bands of 60 to 64 gigahertz and 77 to 81 gigahertz, which inspired the researchers to name their approach “mmSpy.” This is a subset of the radio spectrum used for 5G, the fifth-generation standard for communication systems across the globe.
In the mmSpy demonstration, the researchers simulated people speaking through the earpiece of a smartphone. The brand is irrelevant, Basak said, but the researchers tested their approach on both a Google Pixel 4a and a Samsung Galaxy S20. The phone’s earpiece vibrates from the speech, and that vibration permeates the body of the phone.
“We use the radar to sense this vibration and reconstruct what was said by the person on the other side of the line,” Basak said, noting that their approach works even when the audio is completely inaudible to both humans and microphones nearby. “This isn’t the first time similar vulnerabilities or attack modalities have been found, but this particular aspect – detecting and reconstructing speech from the other side of a smartphone line – was not yet explored.”
The radar sensor data is pre-processed via MATLAB and Python modules, which are computing platform-language interfaces used in this research to remove hardware-related and artifact noise from the data. The researchers then feed that to machine-learning modules trained to classify speech and reconstruct audio. When the radar senses vibrations from one foot (30 cm) away, the processed speech is 83% accurate. That drops the farther the radar moves from the phone, down to 43% accurate at six feet (1.8 m).
Once the speech is reconstructed, the researchers can then filter, enhance or classify keywords as needed, Basak said. The team is continuing to refine their approach to better understand not only how to protect against this security vulnerability, but also how to exploit it for good.
“The methodology that we developed can also be used for sensing vibrations in industrial machinery, smart home systems and building-monitoring systems,” Basak said. “Vibration tracking over time can help assess wear and tear – using our approach could help identify when machinery needs maintenance before it would traditionally be obvious, for example.”
https://techxplore.com/news/
