Source: Academic Headlines
You may, or even most likely, have encountered such a situation:
When you open an app on your phone, the first thing you see is exactly what you were talking about with your friends or family a few minutes ago. For the first time, you'll be surprised. But over time, you will feel accustomed to it and not be strange.
One of the culprits is that microphones are now ubiquitous and always "eavesdropping" on your microphones, which are embedded in mobile phones, TVs, watches and other devices to transmit your voice to neural networks and artificial intelligence systems in real time, helping the recommendation system to make "customized" push services for you.
(Source: Pixabay)
So, how can we avoid such a situation? There are many ways on the Internet, such as playing a completely unrelated song, turning the sound to the maximum, and actively creating some noise; such as not authorizing the long-term opening of the microphone permission, but choosing to take a one-at-a-time authorization approach.
Now, a team of researchers from Columbia University in the United States has proposed a new method: they have developed an artificial intelligence system that only needs to play a very slight sound produced by the system in the room to avoid the occurrence of "monitoring" events.
In other words, this system will disguise the sound of people talking, and will not be heard by monitoring systems such as microphones without affecting normal conversations.
The research paper, titled "Real-Time Neural Voice Camouflage," has been published on the preprint website arXiv.
Moreover, the researchers also said that this artificial intelligence system is easy to deploy on hardware such as computers and mobile phones, and can protect your privacy at all times.
Beat AI with AI
The problem of being utilized by artificial intelligence should be solved by artificial intelligence.
While the team's results from systems such as jamming microphones are theoretically feasible in the field of artificial intelligence, it is still a tough challenge to use it for practical applications fast enough.
The problem is that interfering with the microphone at a given moment to listen to people's conversations may not interfere with the next few seconds of conversation. As people speak, their voices change with the different words and rates they speak, making it nearly impossible for machines to keep up with the fast pace of people talking.
In this study, artificial intelligence algorithms were able to predict the characteristics of what a person would say next, with enough time to generate the appropriate whisper.
(Source: Pixabay)
Carl Vondrick, an assistant professor in Columbia's Department of Computer Science, said the algorithm could prevent rogue microphones from hearing [people's] conversations with 80 percent efficiency, even when people were unaware of information such as the rogue microphone's location.
To do this, the researchers need to design an algorithm that can destroy neural networks in real time, that can be generated continuously as they speak, and that works for most of a language's vocabulary. In previous studies, none of the above requirements could be met at the same time.
The new algorithm uses a signal called "predictive attacks," which can interfere with any word that is trained by an automatic speech recognition model to transcribe. In addition, when attack sounds play in the air, they need to be loud enough to interfere with any rogue "eavesdropping" microphones that may be at a distance. Attack sounds need to travel the same distance as sounds.
The AI system achieves real-time performance by predicting future attacks on signals or words, conditioned on two seconds of input speech.
At the same time, the research team also optimized the attack so that its volume resembles normal background noise, allowing people to converse naturally in the room without being monitored by systems such as microphones.
In addition, they successfully demonstrated the effectiveness of this method in real rooms with natural ambient noise and complex shapes.
But Vondrick also said that the system is only effective for most English words at the moment, and they are applying the algorithm to more languages.
What can we do for ethical AI?