Resume: Researchers have developed a new artificial intelligence algorithm that prevents smart devices like Alexa or Siri from hearing your words correctly 80% of the time. The algorithm is a step towards providing personal agency to protect the privacy of your voice in the presence of smart devices.
Source: Columbia University
Have you ever noticed online ads following you around that look eerily similar to something you recently talked about with your friends and family? Microphones are built into almost everything today, from our phones, watches, and TVs to voice assistants, and they’re always listening to you. Computers are constantly using neural networks and AI to process your speech, in order to obtain information about you. If I wanted to prevent this from happening, how could I do it?
In the past, as shown on the hit TV show “The Americans,” you played music too loud or turned on the bathroom faucet. But what if you didn’t want to constantly scream over music to communicate?
Colombian Engineering researchers have developed a new system that generates silent sounds that you can play in any room, in any situation, to prevent smart devices from spying on you. And it’s easy to implement in hardware like computers and smartphones, giving people the ability to protect the privacy of their voice.
“A key technical challenge in achieving this was getting everything up and running fast enough,” he said. carlvondrickassistant professor of computer’s science.
“Our algorithm, which blocks a rogue microphone from hearing your words correctly 80% of the time, is the fastest and most accurate on our test bench.
“It works even when we don’t know anything about the fake microphone, like its location or even the computer software running on it. Basically, it camouflages a person’s voice over the air, hiding it from these listening systems and not disturbing the conversation between people in the room.”
Stay ahead of conversations
While team results in corrupting automatic speech recognition systems have been theoretically known to be possible in AI for a while, achieving them fast enough to use in practical applications remains a huge bottleneck. The problem has been that a sound that interrupts a person’s speech now, at this specific moment, is not a sound that will interrupt speech a second later.
As people speak, their voices are constantly changing as they say different words and speak very quickly. These disturbances make it nearly impossible for a machine to keep up with the rapid pace of a person’s speech.
“Our algorithm is able to keep up with predicting the characteristics of what a person will say next, giving it enough time to generate the correct whisper,” said Mia Chiquier, lead author of the study and a doctoral student in Vondrick’s lab.
“So far, our method works for most English language vocabulary, and we plan to apply the algorithm to more languages, eventually making the whisper sound completely unnoticeable.”
Launch of “predictive attacks”
The researchers needed to design an algorithm that could break neural networks in real time, that could be generated continuously as you speak, and that would be applicable to most of a language’s vocabulary.
While previous work had successfully addressed at least one of these three requirements, none have achieved all three. Chiquier’s new algorithm uses what she calls “predictive attacks,” a signal that can interrupt any word that automatic speech recognition models are trained to transcribe.
Also, when attack sounds are played over the air, they must be loud enough to disrupt any unauthorized “listening” microphones that might be far away. The attack sound needs to carry the same distance as the voice.
The researchers’ approach achieves real-time performance by predicting a future attack signal, or word, conditioned on two seconds of voice input.
The team optimized the attack to be similar in volume to normal background noise, allowing people in a room to converse naturally and without being successfully monitored by an automatic speech recognition system.
The group successfully demonstrated that their method works within real-world rooms with natural ambient noise and complex scene geometries.
“For many of us in the research community, the ethical concerns of AI technology are an essential topic, but it seems to belong in a separate thought process. It’s like we were so happy that we finally made a driving car but forgot to design a steering wheel and brake,” he says. Jianbo-shiprofessor of computer science and information sciences at the University of Pennsylvania and a leading researcher in machine learning.
“As a community, we need to think ‘consciously’ about the human and societal impact of the AI technology we develop from the earliest research design phase. The study by Mia Chiquier and Carl Vondrick poses the question: ‘how to use AI to protect ourselves against unintended uses of AI?’
“His work makes many of us think in the following direction: ask not what ethical AI can do for us, but what can we do for ethical AI. Once we believe in this direction, ethical research in AI is just as fun and creative.”
Money: This research is based on work supported by NSF CRII Award 1850069 and NSF CAREER Award 2046910. MC is supported by a CAIT Ph.D. Fellowship. The views and conclusions should not be construed as necessarily representing the official policies, expressed or implied, of the sponsors.
About this research news on artificial intelligence and smart devices
Author: holly evarts
Source: Columbia University
Contact: Holly Evarts – Columbia University
Image: The image is in the public domain.
original research: The findings will be presented at the International Congress of Learning Representations