Light Commands: Laser Audio Injection Attacks on Voice-Control Systems
Light Commands: Laser-Based Audio Injection Attacks on Voice-Controllable Systems
Quick read
Based on the paper's full text (AI-generated); refer to the original
The researchers propose the Light Commands attack, which shines an amplitude-modulated laser at the sound port of a MEMS microphone, causing the microphone to treat light as sound, thereby remotely and covertly injecting voice commands into voice control systems. On the 17 devices tested, Amazon Alexa, Apple Siri, Facebook Portal and Google Assistant could all be attacked; a 5 mW laser (equivalent to a laser pointer) was sufficient to control most smart speakers, about 60 mW could control phones and tablets, and the maximum distance reached 110 meters; the attack could also be carried out from another building 75 meters away through a closed double-glazed window. Because devices generally lack user authentication or implement it improperly, attackers could use this to open smart door locks and garage doors, place orders on e-commerce websites, or unlock and start Tesla and Ford vehicles linked to the target Google account. The authors analyzed the physical root cause of microphones' sensitivity to light, discussed software and hardware defenses, and responsibly disclosed to Google, Amazon, Apple, Facebook and other vendors, making it public on November 4, 2019.
- Voice-controllable systems (Alexa, Siri, Portal, Google Assistant) generally lack user authentication, and existing sound-wave injection attacks have limited range (about 7.62 meters) and are weakened by physical barriers such as windows, making remote covert control of physically unreachable devices difficult.
- The authors propose LightCommands: using a laser driver to encode audio signals onto laser intensity via amplitude modulation, shining the laser at the microphone's acoustic port, and exploiting the MEMS microphone's response to light (the ASIC photoelectric effect and the diaphragm photoacoustic effect) to convert light back into sound, thereby injecting wake words and voice commands.
- Validated on 17 devices, commands can be injected at as low as 0.5–60 mW; a 5 mW laser successfully injected commands in a 110-meter corridor, across buildings at 75 meters, and through double-pane glass windows; it also demonstrated using injected commands to open smart locks, garage doors, place shopping orders, and unlock/start Tesla and Ford vehicles linked to Google accounts, and can use infrared lasers and whisper mode to improve stealth.
- The attack relies on line of sight, cannot penetrate opaque obstacles, requires precise aiming, and is harder against mobile devices; the authors propose defense directions such as software-side multi-microphone signal comparison, random questioning, liveness detection, and hardware-side light-blocking barriers, leaving these for future research.
The paper provides measured data on the minimum laser power and maximum attack distance for 17 devices, which teams deploying voice assistants and smart door locks can use to assess physical-layer risk.