Spying Robot Vacuums Really Suck
LidarPhone exploits the LiDAR sensor in robotic vacuums to eavesdrop on your private conversations.
Just when you thought you had all your bases covered for privacy, it turns out that even your seemingly innocent robot vacuum may be spying on you. And all this time you thought it was only collecting dirt from the floor.
A team in Singapore took advantage of the LiDAR sensor typically found onboard robotic vacuums for navigation purposes, and repurposed it to capture sound with an exploit that they call LidarPhone. Laser microphones, in which laser beams reflected off of vibrating objects are converted into audio, are not a new idea. However, laser microphones require sophisticated setups and fine-tuning that is not possible with a stock vacuum cleaner. Getting a vacuum to work as a laser microphone took some clever thinking.
Building on previous exploits, the researchers first worked to control the LiDAR beam position over WiFi, and to wirelessly capture readings from the sensor. With beam position under attacker control, its normal rotation can be slowed enough to capture meaningful vibration data from nearby objects.
While the sensor data is much improved by the exploits, it is still far inferior to the type of high quality data a purpose-built laser microphone produces. To make sense of the data, the team built deep convolutional neural network classifiers. One of the classifiers was designed to recognize the sound of spoken digits, which could reveal sensitive information like bank accounts or social security numbers. A second classifier was trained to recognize the introductory music of various television news programs, which would allow an attacker to infer what one’s political leanings may be. Accuracy exceeded 90% during evaluation of both classifiers.
Digging into the details shows that LidarPhone is quite limited and does not really act like a microphone in the traditional sense. Only predefined audio samples can be recognized, and the model is not generalized. In order to accurately recognize the sound of spoken digits, for example, the model must be trained with samples specifically from the speaker that LidarPhone aims to eavesdrop on.
This is an early prototype device, and it is probable that refinements to the current methods will improve performance in the future. If you have a robotic vacuum, LidarPhone is something you will want to keep your eyes on.