Is the paranoia of being eavesdropped on by technology fuelling your Black Mirror themed nightmares? Worry not, for an artificial intelligence (AI) algorithm has found a workaround.
The new technology, called Neural Voice Camouflage can protect you against being snooped on. Its operation is quite simple – every time you talk, it generates custom audio noise in the background to mask the real audio.
In turn, this confuses whatever artificial intelligence mechanisms may be transcribing your recorded voices in the background. Fascinating, right? This is a classic example of what is called an “adversarial attack.” In essence, it simply throws AI off your trails.
Battle of the AI
Essentially, by using this algorithm, we’re weaponising artificial intelligence against its own kind. Good thing they can’t mutinise (yet)!
While the process sounds straightforward on paper, it is quite complex. As Science.org reported, the machine learning tool first needs to process the entire sound file before knowing how to tweak it. This means that it wouldn’t work when you want to camouflage your voice in real time.
To overcome this handicap, researchers taught a machine learning inspired by neural networks of a brain to (wait-for-it) predict the future! Yep, after training the AI algorithm for hours using recorded speech, the AI learned what a user might say next.
Are humans really that predictable? Regardless, after the right data sets, the AI was taught to process 2-second-long audio clips and to cloud what a person is likely to say next.
Masking audio to protect privacy
Based on what a person says, the AI can predict what is to come next. Based on what was said and the speaker’s voice, the AI then produces sounds that mask whatever is to follow. For instance, “I’m nearly starved myself, for this conquering kingdoms is hard work,” was transcribed as “im mearly starme my scell for threa for this conqernd kindoms as harenar ov the reson.”
The error rate for the automatic speech recognition (ASR) system jumped from 11.3% to 80.2%. “The error rates for speech disguised by white noise and a competing adversarial attack (which, lacking predictive capabilities, masked only what it had just heard with noise played half a second too late) were only 12.8% and 20.5%, respectively,” Science.org wrote. The findings were presented last month at the International Conference on Learning Representations.
Is technology spying on you? New AI could prevent eavesdropping. (2022). Science.Org.