Acoustic microphone locking system

History

2022: System Announcement

In mid-April 2022, scientists at the Columbia Engineering Institute developed a system that generates quiet sounds that can be turned on in any room in order to block smart devices from tracking users. The technology is easy to implement in hardware, such as computers and smartphones, giving people the opportunity to protect the privacy of their voice on their own.

Despite the fact that theoretically the results obtained by the team in the field of distortion of automatic speech recognition systems were already known, their achievements quickly enough for use in practical applications remained the main bottleneck. The problem was that the sound that interrupts a person's speech until April 2022 was not a sound that will interrupt speech a second later. When users speak, their voices are constantly changing, as they utter different words with different speed and tone. These changes make it almost impossible for a machine to keep pace with the rapid pace of human speech.

A system has been released that prevents smartphones from "eavesdropping" conversations

{{quote 'The key technical challenge to achieving this goal was to make it work fast enough! Our algorithm, which manages to block an unauthorized microphone from correctly sensing your words in 80% of cases, is the fastest and most accurate on our test site. It works even when we know nothing about an unauthorized microphone, for example, about its location or even about computer software that works on German. In fact, the algorithm masks a person's voice on the air, hiding it from these listening systems and not interfering with conversation between people in the room, "said associate professor of computer science Carl Vondrick. }} Researchers needed to develop an algorithm that could disrupt neural networks in real time, that could be generated continuously as speech is delivered and would be applicable to most vocabulary words in the language. Although previous work has successfully met at least one of these three requirements, none has reached all three. Lead author of the study and graduate student in the lab, Mia Vondrika, said that the algorithm uses what she calls predictive attacks - a signal that can break any word that automatic speech recognition models are trained to transcribe. In addition, when attack sounds are played on the air, they must be loud enough to disrupt any unauthorized microphone that may be at a great distance. The sound of the attack must be transmitted at the same distance as the voice.

The researchers' approach achieves real-time performance by predicting an attack on a future signal or word based on two seconds of input speech. The group optimized the attack so that it had a similar volume to the usual background noise, which allows people in the room to conduct a conversation naturally and without successful monitoring by an automatic speech recognition system. The group successfully demonstrated that their method works in real premises with natural ambient noise and complex scene geometry.^[1]

Notes